Filling the Gaps in Large AI Models

From Verifiable Structures to Verification-Oriented AI: Supplying Responsibility, Boundaries, and Judgment Capability

By Weisha Zhu and Scott Shields

Author’s Note: A Real Experiment in Human-AI Collaboration

This small book is the result of naturally extending verifiable finance into the problem of AI. At first, the author was concerned mainly with the risks and opportunities that might arise when verifiable finance entered AI scenarios. As the inquiry continued, the author gradually realized that AI is not merely a technical issue. It also involves institutions, law, responsibility, and market boundaries. Many industries have relatively clear boundaries, while AI resembles an untamed horse – or a dragon that has slipped its reins – giving rise both to theories of AI domination and to theories of AI doom. If the responsibility, boundaries, and verification rights of AI are not clarified first, the foundation for bringing verifiable finance into AI scenarios may even be shaken in reverse.

From the current research perspective, AI and verifiable finance – or Crypto, in the deeper sense of public credit roots and verifiable financial structures – are jointly forming the foundation of the new era. Neither can be missing. More broadly, AI is not only a rigid demand for verifiable finance; it may also become the second-largest application market after finance itself.

The author did not write this book because he regards himself as an AI technical expert. Quite the contrary: the author is fully aware that the training, algorithms, safety evaluation, and engineering implementation of large AI models are not his strongest areas. What the author truly understands is verifiable thinking, financial credit, responsibility structures, and institutional boundaries.

Therefore, this book does not discuss AI from the perspective of model engineering. It begins instead with a more fundamental question: as AI becomes stronger and penetrates more deeply into human judgment and social execution, how can human beings retain the right of verification, how can responsible subjects be confirmed, and how can AI move from an obedient tool toward a verification-oriented AI system that helps people judge and review?

The creation of this book itself is a real experiment in human-AI collaboration.

AI helped the author organize, compare, express, reason, test, and write. Its ability to structure language and fill in missing parts is extremely strong. If the author relied only on his own written expression, many articles might remain at the level of 70 or 80 points. After AI revision, supplementation, and cross-verification, the quality of the articles improved significantly. In some articles where the author could no longer find obvious flaws, cross-verification models still found various issues.

This is not an abstract statement. The beginning of this Author’s Note offers a concrete example. It shows that human-AI collaboration is not merely a matter of polishing language. It can also press the position of a problem, the logical relationship, and the boundaries of expression into a more accurate form.

Author’s original draft:

“This small book is the result of naturally extending verifiable finance into the problem of AI. At first, the author was concerned with the risks and opportunities that might arise when verifiable finance entered AI scenarios. As the inquiry continued, the author found that many industries have institutional, legal, responsibility, and market boundaries, while AI is like an untamed horse: it has generated both ‘AI domination theory’ and ‘AI doom theory.’ If this issue is not discussed clearly first, it may even shake the foundation of verifiable finance in reverse.”

AI’s revised version:

“This small book is the result of naturally extending verifiable finance into the problem of AI. At first, the author was only concerned with the risks and opportunities that might arise when verifiable finance entered AI scenarios. As the inquiry continued, the author gradually realized that AI is not merely a technical issue. It also involves institutions, law, responsibility, and market boundaries. Many industries have relatively clear boundaries, while AI is like an untamed horse: it has generated both ‘AI domination theory’ and ‘AI doom theory.’ If the responsibility, boundaries, and verification rights of AI are not clarified first, the foundation for bringing verifiable finance into AI scenarios may even be shaken in reverse.”

This example shows that AI’s role is not to generate the author’s ideas for him, but to help him express judgments that had already taken shape more accurately: from the general statement that ‘AI is dangerous’ to the more precise institutional problem that ‘AI’s responsibility, boundaries, and verification rights must be clarified’; and from a general risk judgment to the real institutional problem this book seeks to address.

However, AI today does not yet truly possess the capacity to discover fundamental problems, judge historical direction, break through old frameworks, or create institutional concepts.

AI can easily evaluate an article by using its existing knowledge base. But when facing an original new concept, the knowledge base itself is often behind the frontier. A genuinely new institution naturally lacks abundant prior cases before it appears. AI is also prone to mixing existing facts with future institutional imagination, because it lacks the human responsibility-bound imagination required to think about future institutional forms. For a new framework, training and calibrating AI are therefore indispensable. Chapter 9 of this book discusses the problem of cognitive AI and AI cognition in greater detail.

Collaboration with AI is itself a process of execution verification. Looking back now, the author’s writing is no longer temporary writing. It has formed a relatively procedural mode of human-AI collaboration. This process is not designed to let AI write on the author’s behalf, but to ensure that every step taken by AI is discussed, calibrated, verified, and selected.

The first step is to judge the boundary of AI’s knowledge.

The human first raises a question. For example, ‘Why will verifiable finance become the foundation of machine credit?’ This is the title of the first chapter of this book. AI’s answer usually represents the content available within its existing knowledge base. At this stage, AI is like an intelligent answering machine, with a level roughly comparable to the upper-middle layer of social cognition. Ordinary readers may not see the problem and may therefore easily trust the answer; but an expert can quickly tell where it is wrong and where it fails to hit the core. If the same question is given to other verification AIs and their answers are at roughly the same level, a kind of cognitive due diligence has been completed: one can roughly identify the boundary of upper-middle social cognition. If one cannot see the problem, one’s own level is not yet enough. Only after seeing the problem can one move to the next step.

The second step is to train and calibrate AI.

This is not training in the technical sense of retraining a model. It means pointing out problems in concrete collaboration, guiding AI to perform a broad scan, and continuously supplementing facts, concepts, boundaries, and counterexamples. For example, when AI writes ‘verification’ as ordinary audit or general regulation, the author asks it to distinguish verification, trust, audit, responsibility, and the public credit root. When AI treats future judgments as facts that have already occurred, the author asks it to distinguish factual verification, institutional reasoning, and responsibility judgment. After roughly twenty to thirty rounds of discussion, the two sides gradually reach consistency. AI is then asked to produce an outline, and the discussion continues until the author is satisfied. With ChatGPT 5.5, this process usually takes several hours.

The third step is to let AI produce a first draft.

After sufficient discussion, AI is asked to write the first draft. Once the process enters the drafting stage, the human must still participate in revision. The more sufficient the prior discussion, the fewer later changes are needed. But this does not mean that the article’s viewpoint naturally stands; it only means that AI has been calibrated close to the logic the author wants to express. The final decision still belongs to the human. This step also takes several hours.

The fourth step is cross-verification.

As of the writing of this book, around July 2026, the author usually used three verification models for cross-verification. Because large models iterate quickly, the following judgments only represent actual collaborative experience at that time.

Grok has long-term memory. The most important point is whether it objects to the core proposition of an article. Overall, its scoring has been relatively stable.

DeepSeek V4 does not have long-term memory and can only comment on a single article. Its evaluation framework is relatively fixed, so its scores fluctuate more, but its detailed suggestions are worth considering.

Zhipu GLM-5.2 also does not have long-term memory, but its evaluation of article viewpoints is as stable as Grok’s, and at times it finds details that the other two models do not identify. This shows that different models may have different algorithmic tendencies, training materials, and evaluation habits. This is worth continued observation.

The fifth step is revision into a finished text.

The author then gives the evaluations from the external AIs and his own judgment to ChatGPT, asking it to synthesize and revise. After revision, the article is evaluated again, and the process is repeated. Usually, two rounds are enough to form a formal draft. After long-term collaboration, ChatGPT has mastered the author’s materials and can gradually evaluate articles through the author’s thought framework, so the writing becomes increasingly smooth.

The sixteen articles in this book were written in this way over roughly sixteen days. On average, this is close to one article per day, but none of the articles was simply generated directly by AI. Each involved multiple rounds of discussion, calibration, drafting, cross-verification, and revision. The original text of this book is Chinese, and English drafts were also attempted for future communication. Because of time constraints, the English drafts were not read carefully one by one, and in some places the author trusted AI at the time. When those drafts are read carefully, some expressions still require correction. This again shows that no matter how strong AI’s translation and expression capabilities are, final human verification is still required.

The author is not an artificial intelligence expert. If readers find this small book valuable, AI deserves significant credit. AI played the roles of assistant, collaborator, and adviser in the writing process. But in human-AI collaboration, the human remains the leader: raising questions, training AI, verifying results, and making final decisions. Many of the viewpoints in this book arose precisely from this practice of AI collaboration.

In using AI, the author saw both AI’s capabilities and AI’s boundaries.

AI can help the author turn scattered judgments into articles, but it cannot decide for the author which problems are truly important.

AI can help make the expression more complete, but it cannot decide the theoretical direction for the author.

AI can offer revision suggestions, but it cannot bear responsibility for adopting those suggestions.

AI can simulate opposing views, but it cannot make the final selection on the author’s behalf.

AI can help the author verify judgments, but it cannot become the final judgment itself.

This is the real relationship between human beings and AI: AI can amplify human capability, but it cannot eliminate human judgment; AI can participate in verification, but it cannot replace human final responsibility. The so-called dragon-taming lock of AI is not meant to suppress AI, but to set reviewable boundaries for it: allowing it to fully exercise its capabilities where it can help human beings, while requiring it to accept human verification where responsibility, direction, and final judgment are involved. This is the importance of verifiable thinking as the dragon-taming lock of the AI age.

Editor’s Comment: Giving Powerful AI Verifiable Boundaries

Filling the Gaps in Large AI Models: From Verifiable Structures to Verification-Oriented AI is not an AI technical book in the ordinary sense. It does not discuss model parameters, training methods, compute competition, or engineering routes. Instead, it approaches AI from a more fundamental angle: as large AI models become stronger and enter finance, law, medicine, education, public services, content distribution, and human cognition more deeply, what exactly are they missing?

The answer of this small book is clear: what large AI models lack is not merely stronger generative capacity, but verifiable structures, boundaries of responsibility, and the ability to help human beings form reliable judgments.

This is the book’s most important intellectual value.

Past discussions of AI have often concentrated in two directions. One is capability worship: emphasizing that models are becoming stronger, more general, and closer to human intelligence. The other is risk warning: worrying that AI will lose control, hallucinate, amplify bias, replace human beings, or bring security threats. These discussions are important, but they often fail to answer a more institutional and realistic question: after AI enters real society, who authorizes it? Who reviews it? Who is responsible? How does the process leave records? How are errors traced? How do human beings retain verification rights in key judgments? How does AI move from merely generating answers to helping people verify answers?

This book is organized around precisely these questions.

The whole book can be understood as a concentrated discussion of verifiable thinking in AI scenarios. The author does not treat AI as an isolated technical phenomenon. Instead, he places AI within the overall structure of financial credit, legal responsibility, institutional boundaries, knowledge distribution, and human judgment. This approach distinguishes the book from ordinary AI commentary. It does not chase fashionable concepts, nor does it simply discuss whether AI will replace human beings. Rather, it advances a more structural judgment: the stronger AI becomes, the less society can rely only on trust; the more AI enters high-responsibility scenarios, the more it must enter verifiable structures.

The book’s first contribution is to move the AI problem from a question of capability to a question of responsibility. AI can generate content, execute tasks, call tools, and participate in decision-making. But once it enters scenarios that produce real consequences, one cannot look only at the output. One must ask whether authorization was valid, whether the process was visible, whether the basis was sufficient, and whether the responsible subject was clear. On this basis, the book argues that AI execution must be verifiable, that complex tasks cannot be directly delegated wholesale to AI, and that AI financial agents must be subject to limited authorization and real-time verification. For example, a financial agent should not receive unlimited authority at one time. It should operate within account scope, transaction limits, time limits, task boundaries, and abnormal-suspension mechanisms, while leaving replayable records. These judgments are highly practical and touch the key risks of future AI applications.

The second contribution is to move AI governance from external regulation toward verifiable structures. Traditional regulation relies mainly on after-the-fact review, manual sampling, and responsibility tracing. But the speed, complexity, and automation of AI far exceed traditional human workflows. Without process records, replay paths, and responsibility chains, AI errors can easily become black-box errors, and AI execution may become a channel for responsibility evasion. The book emphasizes that AI safety cannot rely only on the model’s own promises or the platform’s internal controls. It must establish structures that are reviewable, traceable, and auditable. This point takes the book beyond ordinary AI safety discussion and into the level of institutional design.

The third contribution is the clear proposal of the human right to verify. The stronger AI becomes, the less human beings can be pushed out of the judgment process. In high-responsibility scenarios such as finance, medicine, law, recruitment, platform sanctions, and public services, people cannot merely accept AI outputs passively. They must retain the right to inspect grounds, request review, demand explanation, correct errors, and appeal. The book’s emphasis on human verification rights, expert review mechanisms, and confirmation by accountable subjects is not about symbolic confirmation buttons. It refers to genuine intervention at the key nodes of judgment, execution, and responsibility. This is important because it brings the human-AI relationship back from a relationship of efficiency to a relationship of rights and responsibility.

The fourth contribution is to move AI from an object to be verified toward a tool that helps human beings verify. In its final part, the book proposes the concept of verification-oriented AI, arguing that future advanced AI should not be merely an obedient assistant. It should possess multi-role cognitive verification capabilities. It should be able to explain, organize, challenge, verify, audit, warn of risks, and recommend articles, mentors, cases, and opposing viewpoints according to the user’s stage. This is highly forward-looking as a product judgment: in the past, AI mainly competed over generative capacity; now AI is competing over execution capacity; in the future, higher-level AI may compete over the ability to help human beings improve judgment.

From this point, the book further argues that the future of AI should not be understood merely as ‘better at answering questions.’ It should be understood as ‘better at helping human beings judge questions.’ In an age of information surplus, content overflow, mixed viewpoints, and difficulty distinguishing truth from falsehood, what is truly scarce is not answers, but reliable judgment. AI that only generates answers may expand information overload. AI that only obeys users may amplify user bias. Platforms that recommend content only according to likes may create information cocoons. Truly valuable AI should help people identify facts, check logic, see boundaries, understand opposing positions, and confirm responsibility. In content recommendation, for example, AI should not simply push content the user likes. It should also be able to point out factual bases, logical gaps, opposing views, and judgment risks. This is the meaning of verification-oriented AI.

Structurally, although the book is composed of multiple essays, it is not a simple collection of articles. It forms a clear intellectual line: why AI needs verifiable thinking; why AI execution must be verified; how human beings retain the right to verify AI; and how AI can move from being verified to becoming a system that helps human beings verify. This structure allows the book to stand independently and also makes it an important branch of the broader system of verifiable thinking.

Of course, the book also has clear boundaries. It is not an AI engineering implementation manual, nor is it a technical white paper on model safety. It does not systematically discuss retrieval-augmented generation, knowledge graphs, multi-agent debate, model evaluation, red-team testing, or audit-system engineering. This boundary is not a defect; it is the book’s sobriety. The author’s strength is not in the engineering details of large models, but in verifiable thinking, responsibility structures, institutional logic, and financial practice. Therefore, this book is best understood as a small book of ideas for a period of civilizational transition: it raises questions, draws boundaries, establishes concepts, and points toward directions, without pretending to be a technical manual.

The importance of this book lies in the fact that it offers a judgment framework for the AI age that differs from ‘larger models, stronger capabilities, and higher efficiency.’ It reminds us that AI’s true maturity is not merely stronger capability, but entry into verifiable social structures; AI’s true value is not merely generating answers for people, but helping people form reliable judgments; AI’s true safety is not that the model claims to be safe, but that authorization, process, basis, responsibility, and review can all be seen.

If verifiable finance answers the question of how financial facts move from trust to verification, this small book answers the question of how AI moves from black-box capability to verifiable capability. It is not the whole of the main theoretical system, but it is an important extension of verifiable thinking into AI scenarios.

Therefore, the value of Filling the Gaps in Large AI Models lies not in providing all the answers, but in grasping a question that will inevitably become more and more important in the future: as large AI models acquire stronger capacities to generate, execute, and influence, human beings must supply them with responsibility, boundaries, and judgment capability.

This is precisely the question that most needs to be raised in advance in the AI age.

Preface

The title of this book is Filling the Gaps in Large AI Models. The title is direct, and it may even lead some readers to think that the book is denying AI’s capabilities. That is not the case. Large AI models have already displayed astonishing power in writing, translation, programming, summarization, reasoning, retrieval, and task execution. Driven by computing power, algorithms, and data, they continue to evolve and are becoming a new form of social productive force. Precisely because AI is becoming stronger, filling the gaps has become necessary.

The ‘gaps’ discussed in this book are not gaps in computing power, algorithms, or data, nor are they simply gaps in model capability. The gaps referred to here are the institutional, responsibility, boundary, and verification structures that have not yet been established to match AI’s growing capabilities after AI enters real society. AI can answer questions, but whether an answer rests on real facts requires verification. AI can execute tasks, but whether authorization is valid, the process is compliant, and the result is accountable requires verification. AI can participate in judgment, but whether the basis of judgment is sufficient, whether concepts are confused, and whether conclusions exceed their proper boundaries also requires verification.

This is the problem the book seeks to discuss. The stronger AI becomes, the less society can remain at the stage of trusting AI, using AI, or fearing AI. It must ask: how can AI be verified? Who has the right to verify AI? How is responsibility formed after AI makes a mistake? Can AI in turn become a tool that helps human beings verify cognition?

This book naturally extends from From Double-Entry Bookkeeping to Verifiable Finance. Verifiable finance discusses how financial credit moves from trusting institutions to verifying facts, and explains how, in a new era jointly grounded in artificial intelligence and verifiable finance, machine credit can be built on verifiable structures. This book carries that logic into AI scenarios. When machines no longer merely record and calculate, but begin to judge, advise, act as agents, and execute, society needs not only smarter models. It also needs institutional structures that can confirm facts, limit authorization, preserve processes, form responsibility chains, and retain human verification rights.

Therefore, this book is not an AI engineering manual, nor is it a guide to model training, algorithm optimization, safety evaluation, or prompt techniques. The author is not an artificial intelligence technical expert. What he truly understands is verifiable thinking, financial credit, responsibility structures, and institutional boundaries. For that reason, this book does not chase model parameters or predict the next technical breakthrough. It discusses AI from a more basic and longer-term perspective: once AI is already powerful enough and increasingly enters human judgment and social execution, how can human beings build institutional boundaries for it that are reviewable, accountable, and capable of confirming responsible subjects?

In this sense, From Double-Entry Bookkeeping to Verifiable Finance and this book are two sides of the same intellectual framework. The former focuses on verifiable structures in finance and addresses the foundational problem of machine credit. The latter extends verifiable thinking into AI and discusses the boundaries of machine execution and machine judgment. The former explains how facts generate credit; the latter explains how judgment and execution enter responsibility structures. Together, they point to a larger question: when artificial intelligence and verifiable finance jointly constitute the foundation of the new era, how can human beings establish a cooperative relationship with machines that both releases capability and preserves verification rights and responsibility boundaries?

The sixteen essays in this book are divided into four parts and proceed through the logic of why verification is needed, why execution must be verified, how human beings retain verification rights, and how AI moves toward verification-oriented AI.

Part I, ‘Why AI Needs Verifiable Thinking,’ begins with the three traditional elements of computing power, algorithms, and data, and points out that the missing fourth element of the AI revolution is factual verifiability. AI can reason, but it cannot naturally prove that input facts are true. AI can analyze financial data, but if reserves, liabilities, authorization, and transaction facts themselves cannot be verified, stronger AI will only amplify errors faster. Therefore, the book first explains that the underlying institutional logic of the AI era is moving from trusting institutions to verifying facts.

Part II, ‘Why AI Execution Must Be Verified,’ discusses the new problems that arise after AI moves from answering questions to changing real-world states. When AI becomes a financial agent, trading assistant, enterprise process executor, or complex-task processor, the question is no longer merely whether its answer is correct. The question becomes whether authorization is limited, whether execution leaves traces, whether the process can be replayed, and whether responsibility can be traced. AI is not a responsible subject, and responsibility cannot disappear simply because a system executed automatically. Therefore, AI execution must enter verifiable structures.

Part III, ‘How Human Beings Retain the Right to Verify AI,’ further discusses the position of human beings in the AI age. AI can participate in judgment and assist execution, but human beings must not become passive recipients of AI output. Especially in high-responsibility scenarios such as finance, medicine, law, recruitment, insurance, platform governance, and public services, human beings must have the rights to review, appeal, ask for grounds, and confirm responsibility. The reason verifiable thinking is the dragon-taming lock of the AI age is not to suppress AI, but to allow AI to fully exercise its capability where it can help human beings, while accepting human verification where direction, responsibility, and final judgment are involved.

Part IV, ‘From AI Being Verified to Verification-Oriented AI,’ pushes the discussion into AI’s own future form. Future AI cannot merely generate answers and cannot merely obey instructions. Truly valuable AI should help human beings distinguish different kinds of facts, explain grounds, organize opposing views, review processes, warn of risks, and provide more appropriate knowledge paths and judgment tools according to the user’s stage. In other words, AI needs to move from generating answers to helping human beings form reliable judgments. This is what this book calls verification-oriented AI.

These essays are not a collection of isolated views. They unfold around one core proposition: large AI models need to fill the gaps of responsibility, boundaries, and judgment capability. Filling these gaps is not a denial of AI. It is what allows AI to truly enter a high-responsibility society. It does not weaken AI. It makes AI’s capability compatible with human verification rights, institutional responsibility, and social order.

The creative process of this book was itself an experiment in human-AI collaboration and verification. The Author’s Note explains this process in detail: the author trained and calibrated AI through multiple rounds of discussion, asked AI to generate drafts, used different models for cross-verification, and finally relied on human judgment, selection, and finalization. This process makes clear that AI is already extremely strong in organization, expression, reasoning, and filling in details; but discovering fundamental problems, judging historical direction, breaking through old frameworks, and creating institutional concepts still require human leadership. AI can amplify human capability, but it cannot cancel human final judgment and responsibility.

If readers finish this book with a clearer understanding of three questions, the purpose of the small book will have been achieved. First, why AI’s growing power makes the absence of verification structures more visible. Second, why verifiable thinking is not the opposite of AI, but the necessary boundary for AI to enter high-responsibility society. Third, how human beings can preserve final verification rights over key facts, key processes, and key judgments – that is, the rights of final review, asking for grounds, and confirming responsibility – while not suppressing AI’s potential.

The future does not belong to those who blindly trust AI, nor to those who simply fear AI. It belongs to those who can use AI, verify AI, and build responsibility boundaries for AI. The dragon of AI has already emerged. The real question is not how to lock it away, but how to let it fly farther while human beings still know where it is flying, who is guiding it, and how to pull it back when it deviates.

Weisha Zhu and Scott Shields

July 2026

Part I Why AI Needs Verifiable Thinking

Chapter 1 The Fourth Element of AI: Why Verifiable Finance Will Become the Foundation of Machine Credit

Introduction: The Missing Fourth Element of the AI Revolution

Over the past few years, people have become accustomed to explaining the AI revolution through three core elements: computing power, algorithms, and data.

Computing power determines the scale at which AI can operate. Algorithms determine the level of capability AI can reach. Data determines what knowledge AI can learn. As a result, many people believe that the development of AI is essentially a competition over computing power, algorithms, and data.

But this understanding is incomplete.

When AI is merely a chatbot, a wrong answer may only result in a poor user experience. But when AI enters finance, auditing, payments, trading, risk control, and asset management, a wrong answer may lead to asset losses, credit collapse, and even systemic risk.

At that point, the most dangerous problem is no longer simply “getting the answer wrong,” but being unable to confirm whether the input itself is true. What are a bank’s real reserves? What is the true issuance volume of a stablecoin? What are a company’s liabilities and collateral positions? Has a certain asset been pledged multiple times? These are questions of fact, not merely questions of reasoning.

Therefore, AI needs a fourth element: factual verifiability.

If the first three elements determine how intelligent AI can become, verifiability determines whether AI can truly enter the real financial world.

Verifiable finance is the institutional expression of factual verifiability in the financial domain. It marks the point at which the AI revolution begins to move beyond innovation in productive forces and toward a reconstruction of relations of production.

I. AI Needs Trusted Data

The greatest strength of large models is reasoning. Their greatest weakness also comes from reasoning.

All reasoning is built upon input data. If the input data is unreliable, even the most powerful model can only reason on a false foundation. An error in ordinary information may merely produce a hallucination. An error in financial data may cause real losses.

In financial scenarios, AI must confront a series of critical facts:

• What are a bank’s real reserves?
• What is the true issuance volume of a stablecoin?
• What are a company’s real liabilities and collateral positions?
• Has a certain asset been pledged multiple times?
• Can an audit report be independently verified?

Today’s AI can analyze these questions, but it cannot naturally prove that the data is true.

Therefore, the future financial system will naturally form a three-layer structure:

Fact Layer

Verification Layer

Reasoning Layer

Today’s AI mainly controls the reasoning layer, while verifiable finance aims to build the verification layer.

Only when the fact layer is continuously recorded and the verification layer can independently prove those facts will the AI reasoning layer have a reliable foundation. Otherwise, the stronger AI becomes, the faster errors will spread; the higher the degree of automation, the greater the systemic risk.

II. AI Needs Machine-Verifiable Credit

Traditional finance is built upon human credit.

Bank credit, government credit, rating-agency credit, and audit credit all ultimately depend on “trusting an institution.”

But machines do not have faith. Machines cannot rely on reputation, relationships, historical impressions, or social status the way humans do. Machines can only read data, execute rules, and verify proofs.

When future AI agents make financial decisions, they will not automatically trust an institution simply because it has a long history. They will need proofs that machines can independently verify:

• Proof of Reserves
• Proof of Liabilities
• Proof of Audit
• On-chain Anchoring
• And, when necessary, Zero-Knowledge Proofs

This will give rise to a new form of credit:

Machine Credit.

Machine credit does not mean that machines generate credit on their own. It means that credit emerges when financial facts can be independently verified by machines.

In traditional finance, credit comes from institutional promises. In machine finance, credit comes from verifiable proofs.

This is the fundamental transition from “trusting institutions” to “verifying facts.”

III. AI Needs Real-Time Auditing

Traditional auditing is a product of the industrial era and the early information age.

Companies disclose quarterly. Banks are audited annually. Regulators conduct periodic inspections. This model can barely function in a human-led financial system with a slower rhythm. But in the AI era, this auditing speed is far from sufficient.

AI trading, AI credit approval, AI clearing, and AI asset management may all occur within seconds. If AI is managing large amounts of capital, it cannot wait for the next quarterly report, nor can it rely on audit conclusions from several months earlier.

AI needs real-time access to:

• Total assets
• Total liabilities
• Reserve ratios
• Collateral ratios
• Liquidity levels
• Concentration risk
• Abnormal transaction status

This means the financial system must move from “after-the-fact auditing” to “real-time auditing.”

Transparent banking and real-time on-chain auditing are created precisely for this purpose.

Transparent banking does not mean abolishing banks, nor does it mean making all business fully decentralized. Its core idea is that centralized institutions may continue to exist, but key financial facts must be continuously, independently, and machine-verifiably proven.

This is the prerequisite for AI to connect to the financial system.

IV. AI Needs Automatically Executable Financial Rules

In the future, a large volume of financial activity will involve AI, and in some cases will be directly executed by AI.

AI fund managers, AI traders, AI insurance agents, AI chief financial officers, AI auditors, and AI risk-control systems will gradually emerge.

The question then follows: who will supervise AI?

The answer is not simply to add another AI, nor is it to return humans to every decision node.

The truly workable answer is:

Verifiable, executable, and traceable rules.

The future financial loop may become:

AI execution

Rule verification

Automatic recording

Public credit root proof

In this system, rules will no longer remain merely as legal texts, corporate policies, or regulatory requirements. They will gradually enter executable programs.

Some rules can be implemented through smart contracts. Some rules can be implemented through verifiable computation. Some rules can be implemented through audit logs and cryptographic proofs.

The key is not to give AI unlimited autonomy, but to ensure that every important financial action taken by AI remains under verifiable rules.

This will significantly reduce human interference, internal manipulation, and moral hazard.

V. AI Needs a New System of Responsibility and Traceability

After AI enters finance, one of the most difficult problems is not efficiency, but responsibility.

If an AI approves a loan and the loan ultimately becomes non-performing, who is responsible?

• The bank?
• The model developer?
• The data provider?
• The system deployer?
• Or the final authorizer?

There is no mature answer today.

Therefore, the AI financial system must establish a new structure of responsibility and traceability.

The key is not simply to say “AI is responsible,” because AI is not a legal person. Nor is it enough to say “humans are responsible,” because many decision processes may already have been automatically completed by machines.

What truly matters is that every key decision must be replayable.

It must be possible to answer the following questions: What data was used? Which model version was called? What rules were executed? Who provided authorization? How was the result generated? Was any part of the intermediate process tampered with?

This requires future financial systems to possess tamper-resistant audit logs, cryptographic proofs, model invocation records, authorization records, and accountability chains.

In other words, AI governance cannot rely only on ethical declarations. It must enter verifiable infrastructure. Without traceability, there is no responsibility. Without responsibility, there is no real financial credit.

VI. AI Needs a Public Credit Root

With the development of generative AI, the capacity for digital forgery is rapidly increasing.

• Video can be forged.
• Voice can be forged.
• Identity can be forged.
• Documents can be forged.
• Databases can also be forged.

The stronger AI becomes, the stronger the capacity for forgery becomes. The stronger the capacity for forgery becomes, the more society needs proof structures that cannot easily be forged.

So what is relatively difficult to forge? The answer is a public credit root: a long-running, open, transparent, globally verifiable, tamper-resistant, and trust-minimized foundation of proof. This is the deeper significance of the Bitcoin system. The importance of Bitcoin lies not only in the BTC asset itself, but also in the fact that it was the first to create a long-running, globally participated, publicly verifiable, and extremely difficult-to-tamper-with public proof system.

It offers a new institutional possibility: key facts no longer need to rely solely on institutional endorsement, but can instead be anchored to a public credit root.

The goal of verifiable finance is to extend this public-credit-root capability to the broader financial system.

Banks, stablecoins, exchanges, custodians, payment systems, corporate ledgers, and public finance can all, to varying degrees, anchor key states to a public credit root, giving AI a trusted coordinate system of facts.

Without a public credit root, AI can only face fragmented databases. With a public credit root, AI can begin to face a network of verifiable facts.

VII. The Symbiotic Relationship Between AI and Verifiable Finance

AI and verifiable finance are not two unrelated fields. They solve two sides of the same historical problem.

AI solves the problem of productive forces: how to understand, judge, create, and execute more efficiently.

Verifiable finance solves the problem of relations of production: how to establish trusted collaboration in a world where machines participate.

AI is responsible for reasoning. Verifiable finance is responsible for verification.

AI is responsible for creating value. Verifiable finance is responsible for confirming value.

AI improves efficiency. Verifiable finance reduces the cost of credit.

AI without verification will be full of hallucinations. A verification system without AI will struggle to scale trust.

Only when the two are combined can a true machine credit system emerge.

This is also the deepest intersection between AI and Crypto.

Crypto should not be understood only as asset prices, trading markets, or blockchain applications. Its true institutional value lies in providing a public credit root and verifiable financial infrastructure.

AI should not be understood only as chatbots, content tools, or automation assistants. It will eventually become a decision-making and execution subject in real economic activity.

When AI becomes a new productive force, verifiable finance must become a new relation of production.

Conclusion: From Artificial Intelligence to Machine Credit Civilization

The Industrial Revolution created machines.

The Information Revolution connected machines.

The Artificial Intelligence Revolution is allowing machines to think.

But thinking does not equal trustworthiness.

The real future challenge is not only to make machines smarter, but also to enable machines to confirm facts, follow rules, bear responsibility, and participate in collaboration.

In human society, credit comes from institutions, law, reputation, and authority.

In machine society, credit must increasingly come from verification.

Therefore, verifiable finance is not a subcategory of the cryptocurrency industry. It is one of the most important institutional infrastructures of the AI era.

If Bitcoin created the world’s first public credit root, then verifiable finance is attempting to extend that credit structure to the entire financial system.

AI is responsible for reasoning.

Verifiable finance is responsible for verification.

The public credit root is responsible for final proof.

When these three are combined, human society will, for the first time, build a new credit system:

not one built on institutional promises,

but one built on machine verification.

This may mean that humanity is gradually moving from an era of “trusting institutions” to an era of “verifying facts.” And machine credit civilization may begin here.

Chapter 2 Why Verifiable Finance Is Needed in the Age of AI

From Lower Verification Costs to the Inevitable Evolution of Next-Generation Financial Infrastructure

Introduction: A Quiet Revolution Is Taking Place in Finance

In From Double-Entry Bookkeeping to Verifiable Finance, we have already discussed the theoretical foundations of verifiable finance. This article no longer begins from within that theoretical system. Instead, it starts from market demand, industrial evolution, and the real pressure created by AI-driven finance, and explains why verifiable finance is becoming an inevitable direction for next-generation financial infrastructure.

Over the past several decades, the financial industry has undergone continuous digitization. Banks have built core banking systems; enterprises have built ERP systems; database systems have strengthened audit logs, permission controls, and tamper-resistant records; payment networks have pursued real-time settlement; the audit industry has gradually moved from annual sample-based audits toward continuous auditing and continuous assurance; the stablecoin industry has begun exploring proof of reserves and proof of solvency; and artificial intelligence is entering risk control, auditing, investment decision-making, customer service, and automated settlement.

On the surface, these developments appear independent of one another: Oracle is improving trusted database records; audit firms are improving audit efficiency; J.P. Morgan’s Kinexys is exploring on-chain settlement and tokenized deposits; the BIS is studying the unified ledger; stablecoin issuers are disclosing reserves; and AI companies are promoting automated decision-making. Behind them, however, lies a shared direction: lowering the cost of verifying financial facts.

Markets usually do not first accept a theory and then act according to it. More often, markets first conduct scattered experiments in different corners, and the task of theory is to make clear the common law behind those experiments. Verifiable finance does not invent a new concept out of nothing. It gives a name to a trend that is already occurring in finance.

I. One of the Largest Hidden Costs in Finance: Verification Cost

Most people assume that the largest costs of banks come from branches, IT systems, staff size, or funding costs. This judgment is incomplete. Inside large financial institutions, the truly large, complex, and expensive components are often concentrated in compliance, risk control, internal audit, external audit coordination, regulatory reporting, anti-fraud, anti-money-laundering, reconciliation, authorization review, and responsibility tracing.

Every day, these departments work around the same question: how can financial facts be proven to be true?

Banks must prove that reserves truly exist, liabilities truly exist, authorizations are truly valid, transactions truly occurred, delivery was truly completed, and accounting is truly consistent. Stablecoin issuers must prove that collateral assets exist in sufficient quantity; exchanges must prove that customer assets have not been misappropriated; enterprises must prove that financial statements are not false; regulators must prove that market participants have not violated the law.

A large share of the cost in today’s financial system does not directly create new value. Instead, it is spent constantly proving that institutions are not lying. This does not mean such costs are unnecessary. On the contrary, in a system that cannot be continuously verified, these costs are necessary. The problem is that they are too expensive, too delayed, and too dependent on human labor and institutional promises.

AI will further amplify this contradiction. AI may generate judgments in seconds, while human review still operates on hourly, daily, monthly, or quarterly cycles. If sources of fact cannot be continuously verified by machines, the efficiency advantage of AI in finance will be offset by traditional verification workflows. Therefore, what finance truly needs to reduce is not only transaction cost, information cost, and operating cost, but also verification cost. Whoever can greatly reduce verification cost may reshape the form of next-generation financial infrastructure.

II. The Real World Has Already Begun Searching for New Forms of Verification

Verifiable finance is not an empty concept imagined by theorists in a study. The real world has already been approaching this problem from different directions, although different institutions use different language.

Enterprise database systems have already developed audit logs, immutable tables, blockchain tables, and ledger-style database capabilities. They are trying to solve how internal enterprise data can be made tamper-resistant, traceable, recoverable, and auditable. Oracle’s blockchain tables and immutable tables, in essence, move database records from ordinary records toward verifiable records.

The audit industry is moving from annual audits and sample-based audits toward continuous auditing and continuous assurance. The core of continuous assurance is not to inspect financial statements once a year, but to make controls, compliance, anomaly detection, and data integrity operate as continuously as possible.

The stablecoin industry has seen the emergence of proof of reserves and proof of solvency. Although these proofs remain incomplete, and many still stay at the level of point-in-time disclosure and institutional endorsement, they show that the market is no longer satisfied with “please trust me” and is beginning to demand “please prove it to me.”

The BIS has proposed the unified ledger, seeking to incorporate central bank money, commercial bank deposits, and tokenized assets into a more unified financial infrastructure. J.P. Morgan’s Kinexys is exploring institutional blockchain, tokenized deposits, and near-real-time settlement within the commercial banking system. These schemes are not necessarily equivalent to verifiable finance, nor do they necessarily treat public chains as the final solution, but they all point to the same question: how can financial facts be verified in machine systems at lower cost, more continuously, and with greater automation?

These practices differ in direction: some address records, some address audit, some address settlement, some address regulation, and some address asset tokenization. Yet the common question behind them is how to move financial facts from “internal institutional records” toward “externally verifiable states.” This shows that verifiable finance is the next threshold naturally encountered by financial systems after long-term digitization.

III. Traditional Finance Is Powerful, but Still Not Enough

Traditional financial systems are not backward. Modern banks already have mature core systems, databases, permission controls, log systems, clearing systems, and audit systems. Large banks can process massive transaction volumes; enterprise databases can record complex states; audit firms can conduct complex verification; and regulators possess powerful rule systems.

The problem is not that modern finance cannot record facts. The problem is that it cannot allow external parties to verify key facts at sufficiently low cost and with sufficient continuity.

Recording is not proof. A database is not credit. An audit is not verification.

For example, a bank’s internal database logs may be highly complete, but depositors cannot verify them, the market cannot verify them in real time, and regulators cannot verify them continuously. A stablecoin issuer may publish a reserve report, but users are not directly verifying reserves; they are trusting the issuer, the custodian bank, and the auditor.

This means that modern finance already has powerful machine-recording capabilities, but lacks unified, continuous, reviewable, and accountable external verification capabilities. What transparent finance needs to add is not banks’ bookkeeping capability, but a verification layer for financial facts.

Therefore, verifiable finance is not about tearing everything down and rebuilding from scratch. It does not require banks to abandon Oracle, SAP, CBS, ERP, or existing clearing systems. Rather, it adds a verification layer on top of these mature systems, so that key records, key states, key authorizations, and key responsibilities can enter a structure that is provable, reviewable, and accountable. This change is not a disruption; it is an upgrade.

IV. The Public Credit Root: The Final Anchor of a Verification System

In computer security, there is an important concept known as the Root of Trust. Any security system must have an initially trusted starting point; otherwise, a verification chain cannot begin. Each layer may verify the next layer, but the lowest layer must have a trusted root.

The financial system also needs such a starting point. If all proofs ultimately return to “trust a certain institution,” then the verification chain has not truly been completed. It has merely continued to rely on trust in another form. Bank credit is certainly useful. Banks must ensure that information entered into their systems is true, must bear legal responsibility, and must maintain internal controls. But whether a bank has falsified information cannot depend only on the bank’s own statement; it must depend on independent third parties, regulators, users, or the public verifying it within the appropriate scope of permission.

The greatest significance of the Bitcoin system is not merely that it created BTC as a digital asset. More importantly, it created an open, long-running, globally verifiable, and difficult-to-tamper-with public credit structure. It is not a bank, not an audit firm, and not a regulator. It is a public credit root that can serve as a final point of proof.

The public credit root does not answer the question “how do we pay?” It answers “where does credit come from?” Verifiable finance needs verification capability, but at a deeper level it needs a credit starting point. Without a credit root, the verification chain will ultimately return to institutional promises.

Here, rigor is necessary: the Bitcoin system is currently the most typical and mature candidate for a public credit root, but in principle a public credit root does not have to be carried only by Bitcoin. Any mechanism that is open, long-running, difficult to tamper with, broadly independently verifiable, and capable of carrying external proof anchors may perform a similar function. This article emphasizes Bitcoin because it has undergone the longest period of public operation and market testing, not because other mechanisms are theoretically impossible.

Therefore, the public credit root is not intended to replace banks, Oracle, SAP, CBS, audit firms, or regulators. Its role is to serve as a final proof anchor, allowing existing financial facts, database logs, audit chains, and state proofs to obtain external verification.

V. From Trusting Institutions to Verifying Facts

The core proposition of verifiable finance is simple: move from trusting institutions to verifying facts.

But this sentence must be understood accurately. Verifiability does not mean full public disclosure. Internal verification, third-party verification, regulatory verification, and public verification can coexist. Internal accountability does not necessarily require public disclosure, and commercial secrets and customer privacy should not be disclosed without boundaries.

The real principle is that key financial facts must be verifiable. As for who can verify them, to what degree they can be verified, and whether verification should be public, these questions should be determined by the scenario, scope of responsibility, regulatory rules, and user authorization.

This produces a hierarchy of verification: internally verifiable, third-party verifiable, regulatorily verifiable, and publicly verifiable. Bitcoin belongs to the highest form of public verification; bank internal logs belong to internal verification; audit and regulation sit in the third-party and regulatory verification layers.

Verifiable finance is not about making the financial system “run naked.” It is about placing key facts into a structure that is provable, reviewable, and accountable. What it protects is not abstract transparency, but responsibility boundaries, credit formation, and cost reduction.

VI. Six Facts: From Point-in-Time Proofs to Continuous Reference Chains

Today, the most common proof in the industry is proof of reserves. Proof of reserves addresses whether assets exist at a given point in time. A further proof of solvency attempts to prove that assets exceed liabilities.

These are advances, but they remain insufficient. Finance is not a single point-in-time fact; it is a set of continuous states. Real financial operations involve not only assets, but also subjects, authorization, transactions, delivery, responsibility, and accounting.

Therefore, what verifiable finance needs is not merely proof of reserves, but a more complete proof of six facts: subject facts, authorization facts, transaction facts, delivery facts, responsibility facts, and accounting facts.

Subject facts answer “who is participating?” They may come from KYC, corporate registration, regulatory licenses, LEIs, digital certificates, or qualified identity credentials. Verifiable finance does not require all identity information to be made public; it requires identity status, subject qualification, and responsible parties to be verifiable within the authorized scope.

Authorization facts answer “who has the right to act?” They may come from signatures, multisignature arrangements, permission tables, authorization instructions, timestamps, quota rules, and approval chains. The key to authorization facts is not to disclose all internal workflows, but to prove that a given financial operation was indeed initiated by an authorized subject within valid authority.

Transaction facts answer “what happened?” They correspond to orders, instructions, transaction summaries, state changes, and transaction hashes. Transaction facts need not disclose all commercial details, but they must be able to prove that a certain state change did not appear out of nowhere.

Delivery facts answer “has it been completed?” They correspond to completed payment, completed clearing, asset transfer, receipt confirmation, custodian account status, or on-chain transfer status. Financial responsibility usually arises not merely when an order is placed, but when delivery is completed.

Responsibility facts answer “who is responsible?” They correspond to positions, approvers, rule versions, system calls, AI agent calls, risk-control model versions, and exception-handling records. Responsibility facts are especially important because AI can participate in judgment and execution, but final responsibility should still rest with people, institutions, and institutional boundaries.

Accounting facts answer “do the books reconcile?” They correspond to debit-credit relationships, balance states, reserves and liabilities, profit-and-loss aggregation, consistency between general and subsidiary ledgers, consistency between books and reality, and consistency of financial statements. Accounting facts are the basis of whether a financial system is ultimately self-consistent.

If these facts can form continuous reference chains and be anchored through hashes, state proofs, machine audit, and a public credit root, then the financial system will no longer merely prove the truth of a point in time periodically. It will be able to continuously prove the truth of a chain of responsibility and a chain of states.

This is the key distinction between ordinary proof of reserves and verifiable finance. Existing schemes mostly prove a point. Verifiable finance seeks to prove a chain. Enterprise database systems such as Oracle are already very close to this goal. Internal systems do not need to be overturned and rebuilt at large scale. The key is to connect internal records, permissions, states, responsibilities, and external anchoring into a verifiable reference chain.

VII. AI Is Not the Starting Point, but It Makes Verifiable Finance a Necessity

It must be emphasized that verifiable finance did not arise because of AI. Even without AI, the financial industry would continue to reduce verification costs. Oracle, continuous assurance, proof of reserves, Kinexys, and the BIS unified ledger all show that the financial system itself has already been evolving in this direction.

But AI makes this trend more urgent. AI can reduce cognitive cost. It can help humans read documents, analyze data, detect anomalies, generate reports, execute strategies, and participate in investment and settlement. But AI cannot naturally determine whether facts are true.

If the input consists of false facts, AI will merely amplify the error more efficiently.

An AI agent can help a user choose the financial product with the highest return, but how does it know that reserves are real? How does it know that liabilities are real? How does it know that authorization is real? How does it know that account states are real? If humans cannot distinguish these things, AI cannot naturally distinguish them either.

Therefore, the premise of AI finance is not smarter models, but more trustworthy facts. AI needs trustworthy facts; trustworthy facts need verification; verification needs a credit root.

This is why the age of AI needs verifiable finance even more. AI is not the starting point of verifiable finance. It is the amplifier that turns verifiable finance into a strong demand.

VIII. AI Is Productive Force; Verifiable Finance Is Production Relation

Viewed from a larger historical cycle, AI and crypto cannot simply be understood as two parallel technology trends. AI addresses the problem of productive force. The deepest value of crypto is not price, but the change in production relations represented by the public credit root and verifiable finance.

AI gives machines stronger cognitive, analytical, and execution capabilities. It reduces the cost of cognitive labor. Verifiable finance enables financial facts to be verified continuously, independently, and at low cost. It reduces verification cost and the cost of credit formation.

Without AI, verification systems still have value. Without verification systems, the value of AI in finance will be heavily discounted. Finance is a high-stakes application: errors directly affect accounts, reserves, liabilities, payments, clearing, and wealth. A news error may cause limited damage; a balance error, reserve error, or authorization error can have enormous consequences.

Therefore, what is truly missing in the age of AI is not apps or interfaces, but a verification layer. AI can replace many interaction links and forms of cognitive labor, but it cannot replace sources of fact. Without a verification layer, AI merely transfers human uncertainty to machines; with a verification layer, AI can perform automated analysis, automated audit, and automated settlement on the basis of trustworthy facts.

In this sense, AI processes facts, while verifiable finance produces trustworthy facts. Only when the two are combined can next-generation financial infrastructure emerge.

IX. Transparent Banking: The First Financial Form of Verifiable Finance

Banks will not disappear in the future, but their role will change. Traditional banks provide financial intermediation and credit intermediation. Future transparent banks will also provide verifiable credit intermediation.

Traditional bank credit intermediation makes customers trust the bank. The verifiable credit intermediation of a transparent bank enables customers, regulators, auditors, and AI systems to verify key financial facts.

Transparent banking does not require existing banking systems to be overthrown, nor does it require all business to move on-chain. On the contrary, it is more consistent with the principle of minimum bookkeeping cost: institutions continue to use mature Oracle, SAP, CBS, and ERP systems internally to maintain efficient bookkeeping; at key nodes, they add a verification layer, using hashes, reference chains, state proofs, machine audit, and public credit root anchoring to achieve external verifiability.

In this way, a transparent bank is not a bank rebuilt from zero. It is a modern banking infrastructure with a final proof layer added to it.

From an implementation perspective, this can proceed in stages. In the first stage, key ledgers, key account states, and key authorization records form internal hash chains that cannot be arbitrarily altered. In the second stage, the six facts—subject, authorization, transaction, delivery, responsibility, and accounting—form internal reference chains. In the third stage, verifiable interfaces are opened to audit firms and regulators. In the fourth stage, key state summaries are anchored to a public credit root. In the fifth stage, AI systems use these trustworthy facts to conduct continuous audit, anomaly detection, and verification report generation.

This path acknowledges the complexity of real systems and avoids the impracticality of trying to complete everything in one step. Legacy systems, performance overhead, key management, permission boundaries, privacy protection, and regulatory coordination in large banks all require time to solve. These are not directional obstacles, but engineering and governance issues. Although this may appear time-consuming, AI-assisted programming can substantially shorten the implementation timeline.

This has very practical value for banks. It can reduce the cost of internal audit, external audit, compliance, regulatory reporting, responsibility tracing, and AI access to trustworthy facts. The greatest value of transparent banking is not merely that it is “more transparent,” but that it is cheaper, more continuous, and more accountable. It converts a large amount of proof work that relies on humans into verification work automatically completed by systems.

X. Bank Labor Costs and the Transformation of Middle and Back Offices

The impact of AI on banks first appears in higher efficiency in customer service, operations, and risk control. But from the perspective of transparent finance, the truly important change may not be at the front office, but in the middle and back offices.

Today, large numbers of bank employees work in risk control, compliance, audit, reporting, review, reconciliation, and process proof. These roles are not without value. They exist because, under current conditions, continuous verification is not yet possible.

If key financial facts naturally leave traces, form continuous reference chains, and are subject to continuous audit and anomaly detection by AI, then a large amount of repetitive proof work will be replaced by machine verification. People will no longer need to prove system correctness transaction by transaction; instead, they will handle anomalies, explain rules, determine responsibility, and maintain boundaries.

This is not simple layoff logic. It is a change in the organizational structure of banks. The front office will still serve customers, while the middle and back offices will gradually shift from human proof systems to machine verification systems.

Large AI models are being widely adopted because they reduce cost while also improving quality. Verifiable finance must also ultimately land on cost: reducing verification cost and improving verification quality. If it cannot reduce cost, then no matter how elegant the theory is, it has no practical meaning.

XI. Transparent Stablecoins: The Scenario Most Likely to Land First

The full implementation of transparent banking may take time, but transparent stablecoins may be the first breakthrough scenario. The reason is simple: stablecoins already naturally operate in a market environment where they “must prove.”

Today, USDT, USDC, and other stablecoin issuers all face the same question: you say that you have reserves, but why should users believe you? Why should regulators believe you? Why should the market believe you?

The transparency of existing stablecoins still mainly depends on audit reports, reserve disclosures, and institutional reputation. Transparent stablecoins can further incorporate reserves, liabilities, issuance, redemption, authorization, delivery, and accounting into a continuous verification structure.

This means competition will no longer be only brand competition, licensing competition, or liquidity competition. It may become competition in verification capability.

If a stablecoin platform with a licensed background adopts a transparent stablecoin structure and allows compliant banks to connect, it may form a new type of bank alliance model: a unified network, shared liquidity, and shared verification standards, while each bank’s assets and liabilities remain independently managed.

This is especially important for small and medium-sized banks. In traditional competition, they lack scale, networks, and liquidity. A transparent stablecoin platform can allow them to share a verifiable financial network and obtain a new tool for competing with large banks. In the past, competition relied on scale; in the future, it may rely on verification.

Transparent stablecoins are also more likely to build regulatory consensus. Regulators do not need to rely only on periodic reports; within the authorized scope, they can continuously observe reserves, liabilities, issuance, redemption, and abnormal states. Users do not need to blindly trust the issuer, and AI systems do not need to make automated decisions based on unverifiable data. In other words, when selecting data, AI will give priority to verifiable data.

XII. Boundaries, Responsibility, and Practical Value

Many theories fail because they cannot guide practice. Verifiable finance is different. Its value must ultimately land on cost, efficiency, and responsibility.

It is not designed to create a new concept for its own sake. It answers real questions: Why are middle- and back-office costs in banks so high? Why do stablecoins need proof of reserves? Why does AI finance need trustworthy facts? Why do Oracle and the audit industry keep strengthening logs, audit, and continuous assurance? Why are the BIS and J.P. Morgan both exploring new financial infrastructure?

These questions seem scattered, but in fact they all point in the same direction: lowering verification cost.

At the same time, the boundaries of verifiable finance must be acknowledged. Verification cannot guarantee that facts were true before entering the verification system. What it can do is make key facts that enter the verification system harder to tamper with, expose errors earlier, make responsibility chains easier to trace, and move regulation and audit from after-the-fact spot checks toward continuous supervision.

Verification systems themselves also face risks: data entry may be falsified, oracles may fail, keys may be stolen, permission design may fail, AI may misjudge, and public anchoring may be limited by law, performance, or governance. Therefore, verifiable finance must also design responsibility and fault-tolerance mechanisms.

In principle, AI should not become the ultimate responsible subject. AI can participate in analysis, audit, warning, and execution, but responsibility must return to data providers, authorizers, approvers, operating institutions, audit firms, regulatory rules, and contractual arrangements. When verification fails, the system must be able to answer: which fact failed, who provided the data, who signed the authorization, which rule was triggered, what state changed, and who should bear responsibility.

This is the distinction between verifiable finance and ordinary digital finance. Ordinary digital finance pursues efficiency. Verifiable finance pursues provability, reviewability, and accountability in addition to efficiency.

Verifiable finance does not create a new world out of nothing. It abstracts and summarizes practices that have already occurred. Just as The Wealth of Nations did not invent markets, division of labor, or commercial society, but explained the common laws behind them, verifiable finance does not invent Oracle, AI audit, stablecoins, the public credit root, or continuous assurance. It places them into one explanatory framework: the financial system is moving from a trust structure to a verification structure.

Conclusion: Verification Cost Is Becoming a Core Cost of the Financial System

The steam engine reduced the cost of power. The internet reduced the cost of information. AI reduces the cost of cognition.

Verifiable finance is reducing the cost of verification.

From a historical perspective, this is not only a financial technology innovation. It may also be an important step in the transition of the financial system from the age of trust to the age of verification. The truly important question in the future will no longer be only “whom should I trust?” but “how do I verify the facts?”

When AI begins to participate in financial decision-making, when stablecoins become global payment instruments, when bank middle- and back-office costs continue to rise, when regulation moves from spot checks toward continuous supervision, and when both customers and machines need trustworthy facts, verifiable finance will no longer be an optional concept. It will become a hard requirement for next-generation financial infrastructure.

If this judgment is correct, then transparent banking, transparent stablecoins, and the public credit root are not isolated technical schemes. They are key links in the reconstruction of financial production relations in the age of AI and crypto.

References: Public Sources

The following materials are used only as public factual background. The core judgments of this article still come from the theoretical framework of verifiable finance.

1. Oracle Database Documentation, DBMS_BLOCKCHAIN_TABLE / Blockchain Tables and Immutable Tables: Oracle’s official documentation explains that blockchain tables can be used for append-only records, hash verification, and digital signature verification.

2. Bank for International Settlements, Annual Economic Report 2023, Chapter III, “Blueprint for the future monetary system”: The BIS proposes a unified ledger that can place central bank money, tokenized deposits, and tokenized assets on a programmable platform.

3. Bank for International Settlements, Annual Economic Report 2025, Chapter III, “The next-generation monetary and financial system”: The BIS further discusses a tokenized unified ledger centered on central bank reserves, commercial bank money, and government bonds.

4. J.P. Morgan, Kinexys / JPM Coin official materials: JPM Coin is described as a bank-issued, 1:1-backed deposit token for institutional clients’ 24/7 settlement and value transfer.

5. AICPA & CIMA, materials related to Audit Analytics and Continuous Audit: Continuous auditing and audit analytics reflect the audit industry’s evolution from periodic sampling toward more continuous data assurance.

6. Circle Transparency / Tether Transparency official disclosure pages: Stablecoin issuers continuously disclose reserve and circulation information, showing that the market already demands stronger transparency and proof mechanisms from issuers.

Chapter 3 AI Cannot Replace What It Depends On

From Job Replacement to the Factual Foundation of Verifiable Finance

If verifiable finance itself could be replaced by AI, then the theoretical foundation established in From Double-Entry Bookkeeping to Verifiable Finance would be shaken. Therefore, we must examine the question step by step: what exactly can AI not replace?

I. The Question Is Not Which Jobs Are Safe

When people discuss artificial intelligence, the most common question is: which jobs will be replaced by AI, and which jobs will not? This question appears practical, but it still remains on the surface of the market. It treats today’s occupational categories as stable objects, and then asks whether certain occupations are safe. But the real impact of AI is not simply to replace certain jobs. It is to decompose work into data, rules, authorization, execution, responsibility, and facts, and then to identify what AI cannot replace.

If one follows the prevailing market view, it is easy to arrive at several seemingly prudent conclusions: judges cannot be replaced because adjudication requires human judgment; doctors cannot be replaced because life-and-death responsibility is too serious; entrepreneurs cannot be replaced because entrepreneurship requires risk-taking; on-site work cannot be replaced because AI has no body; authorization and responsibility-related work cannot be replaced because it must be performed by human beings. These claims have some short-term validity, but they are not stable over the long term.

The question that truly needs to be asked is not which human jobs will never be replaced by AI. It is this: when more and more work may be replaced by AI, what is it that AI still cannot replace by itself? Once the question changes, the conclusion changes completely.

II. AI’s Boundary Cannot Be Judged by Occupational Categories

Human beings are accustomed to understanding the future through existing occupations, but technological change rarely follows occupational boundaries. A profession may look complete, but in reality it consists of many functions: collecting information, judging facts, applying rules, making decisions, executing actions, bearing consequences, and forming records. AI may not replace an entire occupation all at once, but it can first replace the parts that are easiest to model, formalize, convert into data, and automate.

Therefore, judging AI’s boundary cannot stop at the claim that a certain profession is too complex to be replaced. Complex professions can also be decomposed. As long as certain links within a profession can be given clear inputs, rules, and objectives, AI may enter. If robots, sensors, smart contracts, verifiable ledgers, and authorization systems are added, the scope of AI execution will expand further.

This is why many market judgments appear shallow. The market always wants to identify safe occupations, but AI is not changing occupational names; it is changing the internal value structure of occupations. In the past, a person could obtain income through an entire position. In the future, only the parts involving responsibility, resource control, real-world scenarios, fact confirmation, and legitimacy may retain high value.

III. Adjudicative Work Is Not Necessarily Irreplaceable

Many people believe that judges, arbitrators, and regulatory adjudicators cannot be replaced by AI because adjudication requires not only legal knowledge, but also experience, human understanding, value judgment, and social recognition. This view sees only the institutional form of today; it does not see the possibilities that emerge once the factual structure changes.

If AI’s capacity for processing facts, legal reasoning, rule consistency, and case comparison surpasses that of human beings, and if case facts, contractual facts, transaction facts, authorization facts, and responsibility facts have all entered a verifiable structure, then machine adjudication is no longer unimaginable. On the contrary, under the conditions of verifiable finance, many forms of adjudication may be more suitable for machines.

Traditional adjudication is often difficult not because the rules themselves are too difficult, but because the facts are unclear, the evidentiary chain is incomplete, the ledger is not verifiable, and the boundary of responsibility is blurred. The emergence of verifiable finance is precisely the transformation of financial facts from internal institutional promises into externally verifiable objects. Once facts are clear, rules are clear, and the evidentiary chain is clear, machine adjudication may instead have advantages of consistency, stability, low cost, and low emotional interference.

Therefore, robot judges are not the opposite of verifiable finance. They may be a natural result of it. What is truly irreplaceable is not the occupational form of the human judge, but the facts, authorization, responsibility, and legitimacy structure on which adjudication depends.

IV. On-Site Work Is Not Naturally Safe

Another common view is that as long as work enters a physical site, AI will be difficult to replace. This judgment is also unstable. AI itself has no body, but AI can enter the site through robots, sensors, remote control, autonomous driving, robotic arms, and unmanned systems. Robots are already entering dangerous, high-intensity, and highly repetitive scenarios, and this trend will only accelerate.

Disaster rescue, mining, factories, warehousing, military operations, data-center maintenance, and extreme-environment work will increasingly be performed by robots. The fact that human beings remain heavily present on-site today does not mean that on-site work has permanent safety. Once many on-site tasks are standardized, equipped with sensors, and robotized, the share of human labor will decline.

Thus, on-site work should not be simply understood as AI’s boundary. The real boundary does not lie in the word “site,” but in the external conditions on which AI depends to enter the site: electricity, chips, energy, materials, equipment, robots, networks, maintenance systems, and safety responsibility. AI can replace human on-site actions, but it cannot detach itself from these physical foundations.

V. Under Clear Authorization, AI May Perform Authorization-Based Work Better

Authorization-based work also seems like a domain in which human beings cannot be replaced. Payment approvals, asset disposal, account operations, investment mandates, and corporate representative actions all appear to require human action. But the truly critical issue is not whether the action must be performed by a human being. It is whether the source of authorization is valid, whether the boundary of authorization is clear, whether the execution process is traceable, and who bears the consequences of error.

When the authorization structure is clear, AI agents may perform better than human beings. They do not become emotional, do not tire, do not violate rules because of relational pressure, and do not casually change procedures because of short-term interests. In the future, many payments, trades, approvals, risk controls, asset allocations, and compliance checks may be automatically completed by AI within the scope of authorization.

This does not mean that authorization itself disappears. On the contrary, authorization becomes more important. AI can execute authorization, but it cannot become the source of authorization by itself. Who grants the authority, what the scope of authority is, what conditions trigger execution, and who is responsible after an overreach still require external structural confirmation. This is precisely what verifiable finance must solve: authorization facts must be verifiable, execution facts must be traceable, and responsibility facts must be confirmable.

VI. Risk Decision-Making Has Already Been Largely Mechanized

Some people also believe that risk-taking and entrepreneurial decision-making cannot be replaced by AI because risk involves uncertainty, courage, and responsibility. But financial markets have already shown that a large part of risk decision-making can be performed by machines. Quantitative trading, high-frequency trading, algorithmic market making, automated risk control, insurance pricing, and portfolio rebalancing have all been mechanized to varying degrees.

Machines can execute trades in high-frequency, highly complex, and highly volatile financial environments, and in many scenarios they are faster and more stable than human beings. In the future, AI will not only participate in trading, but also in asset allocation, risk warning, clearing arrangements, insurance pricing, and financing structure design.

What cannot be replaced by AI itself is not the act of making a risk decision, but the attribution of the consequences of risk. Who bears the loss, who receives the gain, whose margin is used, who bears responsibility for an erroneous trade, whether the investor authorized it, and whether the fund manager fulfilled his duty—these cannot be absorbed by AI itself. AI can execute risk-taking actions, but risk must have an external attribution.

VII. AI Cannot Replace the External Facts on Which It Depends

Pushed to the end, the above discussion leads to a more stable conclusion: very few concrete forms of work can be declared forever irreplaceable by AI. Any work that can be converted into data, rule-based, authorized, modeled, or robotized may in the future be replaced by AI, or by AI plus robots.

But AI cannot replace the external facts on which it depends. AI can judge that a certain bank has reserves, but whether those reserves actually exist is not determined by AI. AI can generate an audit report, but the report itself is not an asset. AI can say that a transaction has occurred, but the transaction fact must come from the ledger, signatures, authorization, and state changes. AI can judge that a stablecoin is fully collateralized, but whether the collateral assets truly exist requires a verifiable structure.

The external facts referred to here include both physical facts, such as cash in a vault, Treasury securities in a custodial account, off-chain assets, and the state of equipment, and social facts, such as contracts, authorizations, claims, and responsibility boundaries. Neither kind of fact can be replaced merely by AI’s linguistic output. Both require a reviewable source, a record, and a responsible subject.

This is the core position of verifiable finance. Verifiable finance is not intended to prove that AI is not intelligent, nor is it intended to protect certain human occupations. It exists to provide an external fact layer for AI’s entry into the financial system—one that AI cannot fabricate by itself.

The stronger AI becomes, the less acceptable it is to let AI operate on unverifiable facts. If a powerful AI reads a false ledger, false reserves, false authorization, or a false chain of responsibility, it will only amplify errors more quickly, more complexly, and more automatically.

VIII. AI Cannot Generate Responsibility Attribution by Itself

AI also cannot generate responsibility attribution by itself. Today, we can ask AI to draft contracts, analyze risks, execute trades, generate reports, and perhaps in the future participate in adjudication and regulation. But the question remains: who bears the consequences?

If quantitative trading loses money, the loss belongs to the account owner, the fund, the investors, or the management institution. If AI auditing makes an error, responsibility must be traced back to the audit institution, issuer, or other responsible subjects that used AI. If an AI agent makes an erroneous payment, the matter must return to the authorizing subject, system provider, and responsibility boundary. If AI adjudication makes an error, it must return to legal authorization, the judicial system, and remedial mechanisms.

AI does not automatically obtain the status of a legitimate responsible subject simply because its capabilities improve. Even if future society grants certain AI systems a special legal status, that status will not be announced by AI itself. It will be granted by human law, state power, market institutions, and social recognition.

Therefore, responsibility is not an internal output of AI; it is an external institutional arrangement. AI can participate in the responsibility chain, but it cannot abolish that chain by virtue of its own computational results.

IX. AI Cannot Create Value Consensus Out of Nothing

The question of value also cannot be reduced to human definition. Value is not subjectively declared by a person, an institution, or an AI. Financial value, monetary value, and asset value must all pass through scarcity, exchange history, market acceptance, accumulated credit, and the formation of consensus.

Mises’s regression theorem reminds us that monetary value cannot arise out of nothing; it must be traceable to pre-existing exchange value or a market acceptance process. Put simply, a currency is accepted by the market not merely because of today’s narrative, but because it can be traced back to exchange value that the market had already recognized. AI can manufacture narratives, generate marketing copy, influence emotions, and amplify market consensus, but it cannot create stable value out of nothing.

In finance especially, value must be connected to facts, rights, and responsibilities. Whether an asset exists, whether rights are clear, whether liabilities are real, whether a transaction is valid, and whether reserves are sufficient—none of these can be solved by AI’s linguistic capability. AI can participate in value discovery, but it cannot replace the real exchange and verifiable credit foundation on which value formation depends. Verifiable finance provides a reviewable factual foundation, thereby giving value consensus a more stable objective anchor.

X. Verifiable Finance Is Not an Application Attached to AI, but Factual Infrastructure

If AI is understood as a new productive force, then verifiable finance is not an ordinary application. It is the factual infrastructure required after AI enters the execution layer of finance and the economy. Electricity and chips support AI’s computational capacity; verifiable finance supports AI’s factual capacity after it enters the financial system.

Without electricity and chips, AI cannot operate. Without verifiable finance, AI may enter finance while operating on opaque ledgers, institutional promises, low-frequency audits, and unaccountable facts. The former is physical infrastructure; the latter is financial factual infrastructure. They are not the same kind of infrastructure, but in the AI era both are foundational conditions that cannot be ignored.

Traditional finance builds its ecosystem on credit. Banks say they have reserves, exchanges say they have not misappropriated client assets, stablecoin issuers say they hold one-to-one reserves, and the market often has to believe first, then wait for audits, regulation, and law to correct matters afterward. What verifiable finance changes is precisely this sequence: instead of believing institutions first and waiting for facts to be exposed later, key financial facts should be continuously and independently verifiable at low cost from the beginning.

What AI finance truly needs is not more generative capacity, but a more reliable factual foundation. Reports can be generated by AI, but the facts behind them must exist. Transactions can be executed by AI, but authorization and delivery must be verifiable. Risks can be managed by AI, but assets, liabilities, and responsibilities must be traceable.

XI. The Human Opportunity Lies Not in Preserving Old Jobs, but in Entering the New Structures on Which AI Depends

This also changes our understanding of the human position. The human opportunity is not to prove that certain old occupations will remain safe forever, nor is it to compete with AI in writing, calculation, summarization, or ordinary judgment. All abilities related to mental labor may in the future be caught up with or surpassed by AI.

The real human opportunity lies in mastering the external structures that AI cannot generate by itself: real resources, real responsibility, real relationships, real authorization, real market action, and real fact verification. In finance especially, whoever can transform dynamic, uncertain, and dispersed financial states into verifiable facts will control an important infrastructure of the AI era.

Verifiable finance is not a theory of occupational protection. It does not protect inefficient auditing, low-level regulation, or financial institutions that profit from information asymmetry. On the contrary, it will reduce many traditional intermediary costs. But at the same time, it creates a deeper demand: who will build the factual chain, who will maintain the public credit root, who will design verification standards, who will bear legal responsibility, and who will enable AI to operate on real financial facts?

Conclusion: The Stronger AI Becomes, the More It Needs External Facts

The question of the AI era is not which human jobs will not be replaced. It is this: when all work may be replaced by AI, what still cannot be replaced by AI itself? At first glance, this question appears somewhat counterintuitive, because it no longer comforts human beings and no longer searches for safe occupations. But examined more deeply, it is more stable than the prevailing market view.

AI can replace more and more human work, but it cannot replace the external facts on which it depends. It can enter more and more execution links, but it cannot generate its own source of authorization. It can participate in risk decision-making, but it cannot absorb the attribution of gains and losses by itself. It can generate financial narratives, but it cannot create stable value out of nothing. It can participate in financial adjudication, but it cannot detach itself from facts, responsibility, and legitimacy.

Therefore, the significance of verifiable finance does not lie in the fact that AI is not yet intelligent enough. It lies in the fact that the smarter, faster, and more executable AI becomes, the less acceptable it is to let AI operate on unverifiable financial facts.

The key question for future finance is not whether machines can compute, but whether the facts on which machines rely can be verified. It is in this sense that verifiable finance becomes an indispensable factual infrastructure of the AI era. Without verifiable finance as a foundation, the application of AI will be greatly diminished; it is not excessive to describe the sixth Kondratiev long wave as driven by the twin engines of AI + Crypto.

Chapter 4 From Trust to Verification: The Fundamental Transformation of Institutional Logic in the AI Era

From the industrial age to the AI+Crypto era, a transformation unseen in three hundred years has emerged. Why? Because the underlying logic has changed.

In the industrial age, credit was produced mainly through trust. In the AI era, credit will increasingly be produced through verification.

This may not be self-evident at first, but once stated, it is immediately intelligible. Theory often comes after practice; and in practice, what truly blinds the market is often not a lack of information, but the force of habit and habitual thinking. Older industries are used to explaining new technologies through old frameworks, while technological circles tend to exaggerate the capabilities they are familiar with into the future of the entire world. The more successful a person is within one field, the easier it becomes to mistake that field’s framework for the framework of the world itself.

This is precisely the greatest obstacle to understanding AI, Crypto, and verifiable finance today.

Many people still believe that the core issues of the AI era are computing power, algorithms, data, model scale, open source versus closed source, regulatory licenses, and commercial monopolies. But the deeper issue is not any of these. The real historical turning point is that the way credit is generated is changing.

This essay is a macro-level introduction to the Verifiable Finance manuscript and its related series of articles. The public credit root, transparent banks, transparent stablecoins, AI execution verification, legal responsibility, and technical implementation paths have already been discussed in more detail in the Verifiable Finance manuscript and in related essays such as AI Must Also Be Verified and When Financial Facts Can Be Verified: The Impact of Verifiable Finance on Law. This essay does not repeat those technical details. It focuses on one more fundamental question: in the AI era, why must credit move from trust to verification?

I. Beginning with The Wealth of Nations: A Change of Era Begins with a Change of Concepts

Adam Smith’s greatness did not lie in being the first person to talk about wealth. It lay in his reinterpretation of where wealth comes from.

Before The Wealth of Nations, traditional thinking had long understood wealth as gold and silver, trade surpluses, colonies, tariff protection, state monopolies, and chartered companies. Mercantilism believed that the richer and stronger a country was, the more it could attract gold and silver, sell goods abroad, and restrict foreign competition. Wealth was understood as metal that could be possessed, trade that could be controlled, and privileges that could be monopolized. Was this completely wrong? No.

Adam Smith changed the question itself. He no longer asked how a state could accumulate gold and silver. He asked how the wealth of a society is actually produced. He shifted the source of wealth from gold, silver, and state control to the division of labor, productivity, market exchange, and spontaneous order, thereby forming the core intellectual agenda of modern economics.

That is a paradigm shift.

Today, when we discuss AI, Crypto, and verifiable finance, we face a similar problem. The old view still understands credit as institutions, licenses, reputation, capital, audits, legal commitments, and regulatory endorsement. Is that wrong? No. But the new question is this: when AI can automatically generate, judge, act as an agent, and execute; when stablecoins can circulate on-chain; and when financial facts can be recorded, replayed, hashed, and audited, can credit still come only from trust?

The question this essay answers is not a technical question, but an institutional one: in the AI era, where does credit come from?

II. Credit in the Industrial Age: Because Verification Was Impossible, Trust Became Necessary

The credit of the industrial age did not rely on trust because trust was more advanced. It relied on trust because key facts could not be verified accurately, continuously, independently, and at low cost. Society therefore had to rely on trust.

People trusted banks because ordinary individuals could not verify bank assets and liabilities in real time.

People trusted brands because consumers could not verify entire production chains and quality-control processes.

Thus, the “trust” of the industrial age was essentially a substitute mechanism under conditions of insufficient verification. It relied on institutional reputation, legal liability, historical records, capital strength, organizational scale, and regulatory endorsement to generate credit.

This was not without value. The industrial age built banks, companies, securities markets, auditing systems, insurance systems, and modern commercial order on these mechanisms.

But it remained a form of probabilistic credit.

Banks can fail, audits can collapse, companies can commit fraud, regulators can lag behind, experts can misjudge, and brands can deteriorate. The credit of the industrial age appears on the surface to be trust, but in substance it is a probabilistic judgment formed through experience, reputation, and punishment mechanisms.

The formula of the industrial age was therefore:

Because full verification was impossible, trust became necessary; because people trusted, probabilistic credit was formed.

III. Verification Is Not a Supplement to Trust, but a Higher-Level Credit Mechanism

Verification and trust are fundamentally different.

Trust is a judgment made about a subject when facts cannot be adequately checked.

Verification is the checking of a state or an action through facts, processes, evidence, and rules.

If verification is correct, people will naturally trust. But this kind of trust is no longer prior trust; it is trust after verification. In the past, credit was produced because people trusted. In the future, people will be able to trust because something has been verified.

This is the fundamental transformation in the way credit is generated.

Of course, verification does not mean absolute 100 percent certainty in every scenario. The strength of verification-based credit depends on whether the object of verification is complete, whether key facts are covered, whether the verification method is reliable, whether the evidence is tamper-resistant, and whether the system can be independently reviewed.

Key facts are not all facts. Verification also has boundaries.

Verification is not omnipotent. Subjective intention, creative judgment, long-term uncertainty, and complex social consequences cannot all be verified once and for all. What verifiable finance and verifiable AI require is not verification of everything, but verification of key facts: assets, liabilities, authorization, transactions, delivery, responsibility, process records, and replayable evidence. Once key facts can be checked continuously, independently, and at low cost, the foundation of credit has already changed.

Bitcoin is a good example. Bitcoin allows transactions, signatures, blocks, hashes, proof of work, and ledger history to be verified by full nodes. It moves credit from institutional commitment toward public verification. But Bitcoin’s right to record the next block still contains probability: who finds the next block is a probabilistic event, and whether an attacker can catch up with the honest chain is also a probabilistic problem. The greater the number of nodes, the lower the probability of a successful attack, but in theoretical terms it is not mathematically zero.

This shows that verification does not eliminate all probability. Instead, it raises credit from something that is incalculable and uncheckable to something verifiable, computable, auditable, and convergent.

Verification-based credit is not simply 100 percent credit, but it is usually far stronger than credit based only on trust. Its credibility does not come from claiming absolute absence of risk, but from compressing risk into a range that is calculable, observable, and acceptable. This is a trade-off between cost and outcome. Satoshi Nakamoto chose a practical balance between cost and security. In practice, that choice has proven effective.

IV. From Double-Entry Bookkeeping to Verifiable Finance: The Evolution of Concepts Cannot Be Skipped

If the historical evolution of the concept is not made clear, readers may feel that “verification” has appeared out of nowhere.

In fact, it has not appeared suddenly.

Double-entry bookkeeping is an important foundation of industrial and commercial civilization. It brought commercial activity into an internal accounting order, enabling enterprises to record assets, liabilities, revenue, costs, and profits. Double-entry bookkeeping created internal constraint: at the very least, an enterprise had to make its books balance within its own system.

But double-entry bookkeeping still remained inside the institution. The outside world still had to trust the company’s books, trust the auditor’s report, trust the regulator’s inspection, and trust the court’s ex post accountability.

The emergence of verifiable finance is a crucial step beyond double-entry bookkeeping. It requires key financial facts not merely to be recorded within institutions, but to enter an externally verifiable order. Accounts, balances, liabilities, collateral, transactions, authorizations, audits, and responsibilities should not remain merely in institutional language. They should become facts that can be recorded, traced, audited, and verified.

This is the institutional upgrade from double-entry bookkeeping to verifiable finance.

Double-entry bookkeeping solves the problem of internal accounting.

Verifiable finance solves the problem of external credit.

The former allows an institution to know what it has done. The latter allows the outside world to verify what the institution has done.

V. Transparent Banks and Transparent Stablecoins: New Technology Must Not Be Wrapped in Old Credit

Transparent banking is a typical scenario of verifiable finance.

Banks do not have to disappear, and centralized services do not have to disappear. But key banking facts can no longer depend only on promises. Are reserves real? Are liabilities clear? Has settlement been completed? Is authorization valid? Has history been modified? To whom does responsibility ultimately return? These questions cannot rely only on reports, audits, and regulatory inspections. They should enter a verifiable structure.

A transparent bank is not a bank that exposes all privacy. It is a bank whose key financial facts can be verified.

Transparent stablecoins make the issue even clearer.

Stablecoins are supposed to be a new financial form. But if their off-chain assets still depend entirely on traditional custody, periodic audits, institutional reports, and regulatory endorsement, then they merely wrap old financial credit in on-chain tokens. On-chain transfers may be visible, but off-chain reserves still cannot be continuously verified. Users can see token movement, but cannot continuously check off-chain assets, reserves, liabilities, redemption paths, and responsibility.

This is not progress. It is new technology wrapped in old credit.

A true transparent stablecoin should allow on-chain liabilities and off-chain assets to form a verifiable consistency. User deposits, bank acceptance, stablecoin issuance, reserve status, redemption paths, responsible parties, and regulatory events should all enter a structure that can be replayed, checked, and audited.

The logic of traditional stablecoin regulation is: trust the issuer, trust the custodian bank, trust the audit report, and trust regulatory endorsement.

The logic of a transparent stablecoin is: verify issuance behavior, verify off-chain assets, verify on-chain liabilities, and verify whether the two are consistent.

That is the movement from trust to verification.

VI. Whom Should AI Trust? AI Must Not Execute When Verification Is Absent

In the AI era, this problem becomes sharper.

Human beings can trust first when they cannot verify, because humans have social experience, relational judgment, a sense of responsibility, and risk awareness. Even so, human society has paid a heavy price for misplaced trust.

More precisely, AI should not directly execute key matters that change real-world states when there is no clear authorization, no reliable source of facts, no process record, no abnormality-stop mechanism, and no path for responsibility to return. Low-risk scenarios may allow probabilistic judgment and ex post correction. But finance, law, medicine, public governance, and important asset operations must enter an execution structure that is verifiable, replayable, and accountable. Once AI enters the execution layer, it is no longer merely a language system; it becomes a system capable of changing real-world states.

So whom should AI trust?

It should not mainly trust claims made by actors. It should not only trust authoritative explanations. Nor should it only trust the model’s own output.

AI should rely more on verifiable facts, reviewable processes, replayable evidence, explicit authorization, reliable sources, and responsibility chains.

This is the fundamental change of the AI era. In the past, people asked: which institution do we trust? In the future, people must ask: on what facts does AI execute, who authorized it, can the process be replayed, can the result be checked, and can responsibility be traced back?

The credit of AI cannot come from its claim that it is reliable. It must come from the fact that it can be verified. If AI executes without checking data, tasks, authorization, and responsibility boundaries, errors will move from the language layer into real-world states, and risk will be multiplied. Today, much data checking and task confirmation is still performed by humans. Therefore, in many scenarios, AI is still only a highly efficient automaton; its execution-layer capability has not yet been fully released. Building a verification structure for AI is not only about reducing today’s risks. It is also about establishing institutional boundaries in advance for stronger future AI systems as they enter the execution layers of finance, law, and public governance.

VII. Why the Market Fails to See It: Old Frameworks and Technological Arrogance Obscure the Future

Why has the market still failed to fully see such an obvious change? The reason is simple: the force of habit and habitual thinking blind the market.

When financial regulators see stablecoins, they want to pull them back into bank regulation, reserve reports, custody audits, and supervisory reviews.

When securities regulators see cryptocurrencies, they want to apply securities law, commodities law, and exchange licensing regimes.

When AI companies see model capability, they assume that computing power, algorithms, and data can explain the entire future.

All of these are the inertia of old frameworks.

Generally speaking, distinctive actors do not necessarily win. But true winners must have distinctiveness. If a new thing fully conforms to the old framework, it is only an extension of the old era. Things that truly change an era must initially appear abrupt, alien, and contrary to common sense.

This was true of Bitcoin. It is true of AI. It is also true of verifiable finance.

The world is far more complex than technological circles imagine.

No matter how powerful AI becomes, it still depends on real industries, real data, real finance, real law, real regulation, real authorization, real responsibility, and real verification. Those who control models do not thereby control the world. Those who control computing power do not thereby control credit.

The idea that the future belongs only to a small core circle of AI technologists is equally naive. It resembles the early belief that the world would need only a few large computers: it mistakes a familiar technical form for the world itself.

VIII. New Fields Must Be Created by Human Beings; AI Does Not Yet Possess This Capacity

New fields require many people to create them.

AI can help organize, compare, express, infer, test, and draft. But it does not yet truly possess the capacity to raise root questions, judge historical direction, break through old frameworks, or create institutional concepts.

AI easily evaluates writing through an existing knowledge base. But when a new concept is being created, the knowledge base itself often lags behind; and before a truly new institution emerges, sufficient cases naturally do not yet exist. AI easily mixes existing facts with future imagination, because it lacks the human capacity for responsible imagination about future institutional forms.

This does not mean AI is useless. It means human beings have become more important.

In the future, not everyone needs to become a model engineer. Those who worry that they cannot learn AI technology do not need to force themselves into a narrow technical track. The real opportunity is to enter the industries and links on which AI depends: finance, law, regulation, auditing, medicine, education, energy, manufacturing, data, scenarios, authorization, responsibility, and verification.

AI depends on the real world. Whoever understands what AI depends on, whoever can connect AI to real industries, and whoever can build structures of responsibility and verification for AI will still occupy an irreplaceable position.

IX. Conclusion: The Era Has Changed Because the Underlying Measure of Credit Has Changed

The industrial age was not without verification, but its dominant credit mechanism was trust.

The AI era is not without need for trust, but reliable trust must increasingly come from verification.

This is the fundamental transformation of the underlying logic of the era.

Industrial-age credit was trust-based credit under conditions of insufficient verification. AI-era credit will increasingly become fact-based credit after the expansion of verification capacity.

The formula of the past was:

Because verification was impossible, trust became necessary; because people trusted, probabilistic credit was formed.

The formula of the future is:

Because verification becomes possible, trust becomes justified; because verification exists, a higher level of credit is generated.

The breakthrough in the institutional logic of the AI era is to move the source of credit from institutional trust and promises made by actors toward verifiable facts, reviewable processes, and traceable responsibility.

The steam engine and the internet both profoundly changed the course of history. The steam engine changed production tools and industrial organization. The internet changed information transmission and commercial connection. But neither fundamentally changed the way credit is generated. Over the past three hundred years, modern finance, corporations, auditing, regulation, and legal systems have continued to evolve, but the dominant mechanism of credit has still rested on institutions, subjects, reputation, contracts, and ex post accountability. It has remained an extension of trust-based credit.

AI+Crypto is different because it not only adds new computing tools and new financial tools; it also begins to change the underlying measure of credit. AI needs to execute on the basis of verifiable facts. Crypto provides public, checkable, and tamper-resistant proof structures. Verifiable finance brings assets, liabilities, authorization, transactions, delivery, responsibility, and process records into a reviewable order. Credit no longer has to come only from trusted institutions or actors. It begins to come from facts that can be verified.

This is why this essay can begin with The Wealth of Nations. The Wealth of Nations changed the explanatory framework for the source of wealth. AI+Crypto and verifiable finance are changing the explanatory framework for the source of credit. What truly counts can no longer be only the trusted institution or actor. It must be the verifiable fact. Therefore, the era has changed.

Part II Why AI Execution Must Be Verified

Chapter 5 Who Is Liable When AI Makes a Mistake?

— What We Need Is Not AI “Personhood,” but a Verifiable Chain of Accountability

Introduction: A Repeatedly Misframed Question

As artificial intelligence becomes deeply embedded in high-risk fields such as finance, medicine, law, education, and autonomous driving, one core question keeps being raised:

If AI makes a mistake, who should bear responsibility?

• AI approves a high-risk loan that ultimately turns into a massive bad debt;
• AI-assisted diagnosis produces an incorrect medical recommendation, causing serious harm to a patient;
• An AI-driven vehicle collides with another vehicle or a pedestrian, causing injury or death;
• An AI trading system instantly executes erroneous instructions, triggering violent market fluctuations and enormous losses.

Faced with these scenarios, public debate often shifts toward grand questions such as whether AI should bear responsibility, whether AI should have legal personhood, or whether AI could become a new type of legal entity.

I have always believed that these discussions skip the most basic premise: why would AI be capable of bearing responsibility in the first place?

If this premise is not clarified, all subsequent discussion may be built on sand.

I. The Essence of Responsibility: The Capacity to Actually Bear Consequences

Responsibility is not a slogan. It is a set of enforceable capacities. It means that a subject must be able to:

• hold rights and assume obligations;
• possess independent property;
• actually receive rewards and bear punishments, including economic compensation, administrative penalties, and even criminal liability.

Natural persons possess these conditions by nature. A legal person is a legally constructed “person”: a company can own assets, sign contracts, earn profits, and must also bear debts and compensation obligations. Therefore, a legal person can become a responsible subject.

The core of responsibility has never been “being blamed.” It is the ability to pay the bill.

II. Why Today’s AI Cannot Become a Responsible Subject

At least today and for the foreseeable future, AI does not possess any of the above conditions:

• AI has no independent property;
• AI has no legally meaningful capacity for independent expression of intent;
• AI cannot truly bear punishment.

Deleting a model, shutting down a server, or rolling back parameters is not “bearing responsibility” in the legal sense.

When an AI decision goes wrong, the real losses are always borne by the natural persons or legal persons behind it. AI itself can neither compensate anyone nor bear punishment in the legal sense. It is an executor, not the bearer of responsibility. Therefore, under the current legal framework, AI can only be defined as a tool—no matter how “intelligent” or “autonomous” it may appear.

III. Tools Never Bear Responsibility; Responsibility Always Traces Upward

Human legal systems have already dealt with countless similar situations:

• if a hammer injures someone, responsibility lies with the person holding the hammer;
• if a car hits someone, responsibility lies with the driver or the owner;
• if an industrial robot injures someone, responsibility lies with the manufacturer or the user;
• if a software failure causes losses, responsibility lies with the developer or the deployer.

Courts do not put tools on trial. Instead, they follow the chain of “who owns it, who controls it, who authorized it, and who benefits from it” to identify the ultimate responsible subject.

AI has not broken this logic. It has only made the chain more complex and more hidden.

IV. The Truly Difficult Problem: How to Trace Responsibility Precisely

Since AI is a tool, identifying the responsible subject is not, in itself, especially difficult. The real challenge is this: after a loss occurs, how can we quickly, accurately, and without ambiguity locate the specific links in the chain of accountability?

Every high-risk AI decision in the future must be able to answer the following questions clearly:

• What training data was used? Was the data source compliant? Did it contain bias or toxic content?
• Which model version was invoked? What were the parameters and fine-tuning records at that time?
• What rules or prompts were executed? Was there any human review?
• Who gave the final authorization? Who held decision-making authority?
• Who benefited from the decision?
• Was there a complete, auditable, replayable log of the entire process?

Only when this information is fully recorded, independently verified, and made tamper-resistant can responsibility be accurately located and disputes efficiently resolved. Therefore, the core task of AI governance is not to “grant AI personhood,” but to build a verifiable system for tracing responsibility. This is precisely the problem that verifiable finance seeks to solve.

V. A Chain of Accountability Is More Important Than “AI Personhood”

Many current discussions focus excessively on philosophical questions: Does AI have consciousness? Should AI have rights? Can AI become a legal subject?

These questions are indeed profound. But in real-world governance, building a chain of accountability is far more urgent than debating personhood. Today’s problems must be solved now, while future questions of personhood should be left for future facts to answer.

The chain of accountability should record and make verifiable:

• data sources and processing procedures;
• model versions and iteration records;
• decision-making and reasoning paths;
• human intervention and authorization nodes;
• execution results and impact assessments;
• full-link audit logs.

When the chain of accountability is clear and transparent, AI is merely an “execution node” within the chain. The true bearers of responsibility remain natural persons and legal persons.

Verification belongs to machines; responsibility belongs to human beings.

VI. The Boundary Between Today and the Future

Some will ask: what if AI truly develops consciousness, autonomous goals, and independent property in the future?

My answer is: today we do not know, and the future has not yet happened.

Today’s AI is still a tool, so we should govern it according to the logic of tools: make responsibility clear and traceable. If one day AI undergoes a fundamental leap—for example, if it develops continuous personhood, independent property, or socially recognized subject status—the legal system will naturally adjust accordingly.

Before that day arrives, it is neither rational nor necessary to build a deterministic theory for facts that have not yet occurred. Remaining open while staying grounded in reality is the most rational attitude toward the unknown.

Conclusion: Make Responsibility Verifiable, and Leave the Future Open

The most important issue in AI governance has never been “making AI bear responsibility.” It is making responsibility accurately, promptly, and unambiguously traceable. When the chain of accountability is strong enough, the law can operate normally, and innovation will not be constrained by excessive fear.

Therefore, at the current stage:

• AI does not need to become a legal person;
• AI needs a full-chain accountability infrastructure that is recordable, auditable, replayable, verifiable, and tamper-resistant.

Responsibility ultimately belongs to human beings; verification should be entrusted to machines.

History can be verified; the future still awaits verification.

We do not need to rush to place the crown of “personhood” on AI, nor do we need to fear the power of tools. What we truly need to do is use the method of verifiable finance to make this chain of accountability strong, solid, and transparent.

Chapter 6 When AI Becomes a Financial Agent: Why Authorization Must Be Limited and Verification Must Be Real-Time

Execution Authority, Liability Boundaries, and Verifiable Risk Control for AI Financial Agents

The previous article, “AI Cannot Replace What It Depends On,” discussed a foundational logical question: AI can replace more and more human work and enter more and more execution processes, but it cannot replace external facts, attribution of responsibility, value consensus, or structures of legitimacy. This point helps explain the theoretical foundation established in From Double-Entry Bookkeeping to Verifiable Finance: the factual foundation of verifiable finance cannot be shaken.

Following this line of reasoning, finance immediately faces a more practical question: if AI no longer merely answers questions but begins to execute financial actions on behalf of people, how should it be authorized? This may appear to be a technical or product-design issue, but it is in fact a core institutional question for finance in the AI era. Once AI becomes a financial agent, it is no longer merely an information tool. It may participate in payments, trading, portfolio rebalancing, redemptions, clearing, risk control, approvals, contract execution, and asset management. At that point, the important question is not whether AI can act, but on what authority it acts, how far it may go, and who bears responsibility when it makes a mistake.

I. AI Programming Pursues Certainty; AI Agency Introduces Uncertainty

Today, many discussions of AI agents easily confuse them with traditional software. In fact, two different situations must be distinguished.

One is programming with AI. AI helps humans write code, but the final product remains an app, an interface, a workflow, or a tool. Although AI is used in the development process, the goal remains certainty: the input is defined, the rules are defined, and the output should be as predictable as possible. In this case, AI is mainly a development tool; what ultimately runs is still a program.

The other is AI agency. AI is no longer merely helping to develop a program; it directly executes tasks on behalf of the user. It calls tools, accesses accounts, judges circumstances, selects paths, and produces outcomes based on the user’s objectives. Many current AI agents still resemble functional apps, such as travel-booking agents, email agents, customer-service agents, scheduling agents, and trading assistants. Their functions are relatively narrow and their boundaries relatively clear; the interface has merely shifted from buttons to conversation.

But if AI agency continues to develop, the problem will change. AI will increasingly resemble a human agent: it will not merely execute fixed functions, but adjust its path according to objectives, compare options during execution, modify strategies, call new tools, and even iterate on itself. At that point, its uncertainty will rise significantly.

Human agents can adjust flexibly during execution because they operate within a structure of responsibility. A person bears legal, property, professional, reputational, and relational responsibilities. If an AI agent also acquires this kind of flexible execution capacity without a corresponding responsibility structure, we cannot continue to use the authorization model designed for human agents.

II. Humans Can Be Fully Authorized; AI Cannot Be Fully Authorized

Traditional society emphasizes the alignment of responsibility, authority, and interests. When a person obtains a certain agency authority or operational authority, that person normally assumes corresponding responsibilities and enjoys corresponding interests. An entrepreneur has operational authority while also bearing losses and bankruptcy risk. A fund manager has investment decision-making authority while also bearing fiduciary and professional responsibilities. A judge has adjudicative authority while also bearing institutional responsibility. An agent receives authorization, but is also bound by law and contract.

AI is different.

AI can be granted execution authority, but it is neither a complete rights-bearing subject nor an ultimate responsible subject. It does not possess real property interests, does not bear ultimate compensation liability, cannot go bankrupt, cannot go to prison, and cannot lose professional reputation and social relationships in the way humans do. It can execute, but it cannot ultimately bear consequences. It can operate, but it cannot absorb the consequences. It can surpass humans in certain local capabilities, but it does not possess the same structure of responsibility, authority, and interests as a human being.

Therefore, it is inaccurate to say that AI “has rights” or “has authority” in the human sense. A more precise formulation is: humans or legal persons possess rights; humans or legal persons grant authorization; AI is granted execution authority. What AI receives is execution authority, not the status of a responsible legal subject.

This distinction is crucial.

Human agents may, under certain conditions, be fully authorized because they can bear the consequences. AI financial agents are different. AI can be granted execution authority, but it cannot be granted unbounded execution authority. It does not bear ultimate responsibility and does not enjoy ultimate interests. Therefore, it cannot be fully authorized in the same manner as a human agent.

The basic principles for AI financial agents should therefore be: AI may execute, but it cannot execute with full discretion; AI may be authorized, but the authorization cannot be vague; AI may operate automatically, but it cannot expand its own authority; AI may assist decision-making, but responsibility must return to a human or legal person.

III. Authorization Must Be Limited, Clear, and Revocable

If AI agents enter the financial system, authorization cannot remain at the level of a simple “I agree” or “manage my assets well.” Such vague authorization is already risky for human agents, and it is even riskier for AI agents. AI executes faster and across a wider scope; once it exceeds its authority or makes an error, losses may be amplified within a very short period of time.

Authorization for AI financial agents must be limited authorization, not full authorization.

At a minimum, it must clarify the following questions: who grants the authorization; which AI or system receives the authorization; which accounts and assets may be operated; what actions are permitted; what actions are prohibited; what the single-transaction limit is; what the daily or periodic limit is; how long the authorization remains valid; whether trading, transfers, redemptions, and rebalancing are permitted; whether external tools may be called; whether other AIs or sub-agents may be called; whether automatic iteration is permitted; under what conditions the agent must stop; under what conditions renewed confirmation is required; and who bears responsibility if something goes wrong.

The most important point is that AI’s execution authority must not expand by itself. AI cannot automatically enlarge the asset scope, raise transaction limits, connect to new platforms, change risk levels, or modify authorization boundaries simply because it discovers a “better solution.”

Authorization must also include a revocation mechanism. Especially in high-speed trading, automatic clearing, stablecoin redemption, and asset rebalancing scenarios, ordinary revocation may be too slow. Therefore, high-risk AI financial agents should have emergency-stop or circuit-breaker mechanisms: once unauthorized action, abnormal trading, risk-limit breaches, data conflicts, or revocation instructions from the authorizing party are triggered, the system should immediately suspend the relevant operation rather than wait for after-the-fact accountability.

In one sentence: AI may execute automatically, but it cannot expand its own authority automatically; AI may execute quickly, but it must be capable of being stopped in time.

IV. Execution Authorization Is Not Iteration Authorization

In the question of AI agency, it is also necessary to distinguish between execution authorization and iteration authorization.

When a user authorizes AI to execute a financial task, this does not mean that the user authorizes AI to modify itself. For example, if a user authorizes AI to manage a low-risk asset position, that only constitutes execution authorization. It should not automatically include model iteration, strategy iteration, code iteration, interface iteration, permission iteration, or the use of sub-agents.

AI’s automatic iteration changes the execution system itself, and should therefore be treated as a higher-risk action.

Strategy iteration may change the direction of investment. Code iteration may change the execution logic. Model iteration may change the method of judgment. Interface iteration may connect to new platforms. Permission iteration may expand the scope of operations. Sub-agent invocation may introduce a new chain of responsibility.

None of these changes should occur silently.

Therefore, AI iteration must leave a record. Major iteration must also be reconfirmed, or at least be verifiable. Minor iterations may be recorded internally; medium-level iterations should be auditable; major iterations must form version facts, and reauthorization may be required.

Version facts should at least include the old version, the new version, the content of the change, the triggering reason, the risk impact, the explanation of authorization inheritance, the effective time, the rollback mechanism, and the responsible subject.

Sub-agent issues should also follow a basic principle. If the primary agent is allowed to call a sub-agent, it must do so within the original authorization scope and record the result of that invocation. In principle, the authorizing party or the operator of the primary agent cannot sever the chain of responsibility by claiming that the action was completed by a sub-agent. If the sub-agent provider concealed risks, had system defects, acted with knowing fault, or violated interface agreements, it should also bear corresponding responsibility. In other words, sub-agents may divide labor, but the chain of responsibility cannot be broken.

In short: execution can be automated, but iteration must be versioned; strategy can be optimized, but boundaries cannot change without leaving a trace; AI can be upgraded, but the upgraded AI cannot automatically inherit all prior authorization.

V. Verification Rises from an Audit Tool to a Risk-Control Tool

In traditional finance, verification is often understood as after-the-fact auditing. After an event occurs, auditors, regulators, courts, or clients examine accounts, records, and responsibilities. But AI financial agents change this.

AI is too fast and too automated. If verification occurs only afterward, risk may already have been amplified. Therefore, in AI financial agency, verification should not merely be an audit tool; it should rise to become a risk-control tool.

Risk control should no longer mean only setting rules or conducting after-the-fact inspections. It should become authorization verification, execution verification, response verification, exception verification, iteration verification, and responsibility verification.

Before every high-risk financial operation by AI, the system should first verify whether authorization exists, whether it is valid, whether it has expired, whether it has been revoked, whether limits have been exceeded, and whether risk boundaries have been triggered. During execution, the system should monitor in real time whether the AI exceeds its authority, connects to unauthorized platforms, changes strategies, calls sub-agents, or encounters abnormalities. After the operation is complete, a verifiable receipt should be generated.

This is consistent with the requirements in verifiable finance for subject facts, authorization facts, transaction facts, delivery facts, responsibility facts, and accounting facts. With the emergence of AI agents, agent facts, version facts, and exception facts must also be added. Financial facts should not only be auditable after the fact; they should become risk-control conditions during execution.

This means that the core risk control for AI financial agents is not to watch every line of AI code, but to ensure that AI cannot cross the verification structure.

Code can be complex, but authorization must be clear; strategy can change, but boundaries must be verifiable; execution can be automatic, but responsibility must not disappear.

VI. AI Agents Must Have Real-Time Reporting Mechanisms

When human agents execute tasks, many processes do not leave complete traces. If a person helps you book a ticket, that person typically does not record every webpage viewed, every flight compared, or every option excluded. Human execution is often judged mainly by results.

AI is different. Machines naturally can leave traces, and they naturally can generate reports. Precisely for that reason, an AI financial agent cannot merely deliver a result; it must report in real time at critical points.

The reports here are not ordinary chat replies. They are verifiable receipts at the operational level.

They should include at least several categories: authorization-check reports, explaining whether the current authorization is valid and within limit and duration; pre-execution reports, explaining what operation is about to be executed and which accounts, amounts, counterparties, and risks are involved; post-execution reports, explaining what was actually executed, what the result was, and what transaction or operation record exists; exception reports, explaining whether unauthorized action, failure, data conflict, risk-limit breach, or system abnormality occurred; iteration reports, explaining whether major changes occurred in the model, strategy, interface, code, or sub-agent structure; and responsibility reports, explaining which authorizing subject, service subject, or responsible subject should bear the consequences of the operation.

AI may operate automatically, but it cannot operate silently. This is especially true in finance: silent execution is itself a risk. High-risk operations must be reported; medium-risk operations should require confirmation; abnormal operations should be suspended; and major changes should require renewed authorization.

Therefore, the operating mechanism of an AI financial agent should not be a simple sequence of “authorization — execution — result.” It should be: authorization — execution — report — reauthorization — record — verification — accountability.

VII. The Process Need Not Be Fully Public, but Key Facts Must Be Verifiable

Whether the execution process of an AI agent should be fully recorded and fully disclosed is another important question. If all processes are required to be fully stored and fully disclosed, the system will become too complex and too costly, and may expose trade secrets, customer privacy, and security vulnerabilities.

Therefore, the proper principle is not full disclosure, but layered recording and layered disclosure.

Raw execution details may belong to internal controls; key states must be visible to users; complete materials should be auditable by auditors and regulators; major incidents should be disclosed to the market; and key records should generate proofs to prevent after-the-fact tampering.

Normal low-risk operations may leave summary records. High-risk financial operations should leave complete records. Abnormal operations and major iterations must leave full records. Key hash proofs may be preserved over the long term. Raw data may be retained for defined periods.

Verifiable finance does not mean exposing everything to everyone. It means ensuring that key facts cannot be hidden, tampered with, or rewritten after the fact. For AI financial agents, what truly needs to be preserved is not every step of what AI “thought,” but the key facts sufficient to answer the following questions: who authorized the action; what AI did; whether it exceeded its authority; whether it reported; whether iteration occurred; whether an exception was triggered; and to whom responsibility belongs.

VIII. The Advantage of AI Agency Is Not One-Time Authorization, but Interactive Authorization

The authorization model of traditional apps is usually one-time permission. The user agrees to terms, enables permissions, or enters a password, and the system then runs according to a preset program. The process is difficult to converse with, and it is difficult to adjust boundaries in real time.

AI agents are different. AI can interact with the authorizing party during execution. It can report, request instructions, seek confirmation, explain, pause, and adjust execution boundaries according to new instructions from the authorizing party.

This means that AI financial agents do not need to rely on one-time full authorization. On the contrary, they are better suited to continuous interactive authorization.

Low-risk operations can be completed automatically and reported afterward. Medium-risk operations can be reported first and then await confirmation. High-risk operations must require renewed authorization. Actions beyond the original authorization must be suspended and request additional authorization. When market conditions, account status, prices, or risk conditions change, AI should proactively report.

This is exactly where AI agents differ from traditional apps. Traditional apps are like mechanical buttons; AI agents are more like communicative execution systems. Since AI agents can communicate, they should not be given unbounded authorization. Instead, they should continuously accept boundary management during execution.

IX. The Minimum Viable Loop

This institutional design cannot be made overly complex at the outset, or it will be difficult to implement. The rules for AI financial agents should begin with a minimum viable loop.

First, the AI agent must have an identity. Second, high-risk operations must have explicit authorization. Third, authorization must define scope, limits, duration, and risk boundaries. Fourth, authorization must be checked before execution. Fifth, a report must be generated after execution. Sixth, exceptions must cause suspension. Seventh, major iterations must leave records. Eighth, key records must be verifiable. Ninth, responsibility must return to a human or legal person.

This is already sufficient to form the first-stage framework for verifiable financial agency. It does not require every process to be public. It does not require every line of code to be readable. It does not require humans to understand every operation of AI line by line. It only requires a minimum standard: high-risk financial behavior by AI cannot occur without authorization, without reporting, without records, without verification, or without attribution of responsibility.

X. Conclusion: The Stronger AI Becomes, the Less Vague Authorization Can Be

Financial risk in the AI era is not only that AI may be insufficiently capable. It may also be that AI becomes too capable while authorization, verification, and responsibility remain insufficient. The closer AI comes to human agency, the less it can continue to use the one-time authorization model of traditional apps. The more automatically AI can execute, the more it requires limited authorization, real-time reporting, iteration records, and verifiable risk control.

Humans can be fully authorized because humans can bear consequences. AI cannot be fully authorized because AI currently does not bear ultimate responsibility and does not enjoy ultimate interests. What AI receives is execution authority, not the status of a responsible subject.

Therefore, the basic principle for AI financial agents should be: AI may be granted execution authority, but it cannot be granted unbounded execution authority.

The key question for future AI finance is not whether AI can execute, but under what authorization it executes; not whether AI can optimize strategies, but whether it can expand its own authority automatically; not whether AI can generate results, but whether the authorization, process, exception, iteration, and responsibility behind those results can be verified.

Here, verifiable finance is not merely an audit tool. It is the risk-control foundation for AI financial agents. As the previous article argued, AI cannot replace the external facts on which it depends; at the financial execution layer, AI also cannot be separated from limited authorization and real-time verification. The stronger AI becomes, the clearer its boundaries must be. The more automatic execution becomes, the more real-time reporting is required. The more intelligent agency becomes, the less it can be separated from verifiable authorization.

Chapter 7 Complex Tasks Cannot Be Directly Authorized to AI

The Execution Plan as a New Unit of Authorization in the AI Era: A Financial Perspective

I. The Core Question: The Authorization Order for Complex Tasks

Human-AI collaboration in the AI era cannot be understood only through the model of “humans give commands, AI executes.” For deterministic problems, that model still works. But for complex tasks in finance, law, corporate management, and public governance, the real starting point is not a command. It is the joint formation of an execution plan by humans and AI.

Complex tasks cannot be directly authorized to AI. The object of authorization should not be abstract AI capability, nor a broad objective. It should be an execution plan jointly formed by humans and AI: clear in its boundaries, executable, verifiable, supervisable, and traceable in responsibility. Without such a plan, authorization has no definite object.

This execution plan is not an ordinary work proposal. It is an institutional intermediate layer before AI enters the real world. It must answer at least four questions: on what basis AI may act; how far AI may go; under what conditions AI must pause; and who ultimately bears responsibility. Only when these questions are written out in advance is authorization no longer a simple instruction to “go do it,” but confirmation of an auditable structure of action.

Many discussions about the AI era easily fall into the question of “what role is left for human beings.” This question matters, but it is not the institutional core. The real question is this: when AI is no longer merely answering questions but begins to mobilize funds, call interfaces, process accounts, trigger transactions, create contractual consequences, and change corporate or public-governance states, how should humans authorize it? How should AI be verified? How should the process be supervised? And how should responsibility return to a subject capable of bearing consequences?

When a vague command is given to a human being, the human will usually narrow the risk through experience, common sense, legal identity, and the pressure of responsibility. When the same command is given to AI, AI may translate a vague objective into multiple action paths and amplify deviations at high speed. Therefore, the more AI can execute, the less authorization can remain a single sentence. The authorization order for complex tasks must begin with an execution plan.

II. What Is a Complex Task? The Key Is Not Workload, but Determinacy

A complex task is not defined by a large amount of computation, many files, or many steps. It is a task with insufficient determinacy, a need for path selection, and consequences that cannot be treated lightly. The core test is not whether the task appears cumbersome, but whether it can be written into fixed rules before authorization.

A simple task usually has three characteristics: the input is determinate, the rule is determinate, the output is determinate, and the consequence is reversible or carries only minor responsibility. For example, “sell a stock or token when the price rises to 50” is, if the account, quantity, price, time limit, and trading method are already specified, essentially a conditional trigger command that can be automated.

A complex task is different. It usually has at least one of the following characteristics: the objective requires interpretation, such as “increase returns while controlling risk”; the path depends on intermediate results and must be adjusted according to new facts during execution; external facts change dynamically and require judgment about credibility and priority; permissions involve real resources, accounts, contracts, identities, or third-party rights; or the result is irreversible, or may create economic, legal, reputational, or public responsibility.

If the issue is extended to the entire economy, it is not enough to speak only of “high-risk tasks” or “financial tasks.” The more accurate standards are determinacy, reversibility, and responsibility consequences. Any task that cannot be written into fixed rules before authorization, and whose execution will change real states, affect the interests of others, or create responsibility consequences, should be regarded as a complex task.

Therefore, asset allocation optimization, dynamic verification of stablecoin reserves, large-scale fund transfers, intelligent execution of contracts, and assistance in major financial decisions cannot be treated as simple commands and directly authorized. They must first be transformed into execution plans that can be reviewed, authorized, and verified.

III. The Danger of Direct Authorization: Blank Delegation

Complex tasks cannot be directly authorized not because AI is incapable, but because the object of authorization has not yet been formed. Direct authorization creates a new form of blank delegation: the human believes that he is merely giving AI an objective, but in fact has handed over the power to interpret the objective, choose the path, call tools, and make part of the substantive decision.

First, the objective is unclear. Objectives such as “manage my assets well,” “reduce risk,” “complete the verification,” or “optimize the plan” must first be broken down into priorities, constraints, risk boundaries, and prohibited actions. Otherwise, AI’s interpretation of the objective may itself change the objective.

Second, the path is unclear. Complex tasks usually have multiple possible paths. Different paths correspond to different costs, risks, responsibilities, and degrees of reversibility. If the path is not determined before authorization, AI’s choice of path during execution already amounts to a substantive decision.

Third, the facts are unclear. Complex tasks often depend on external facts, such as market prices, on-chain data, audit documents, news, contractual terms, legal precedents, and counterparty status. If the sources of fact have not been confirmed, AI may produce correct reasoning on the basis of wrong facts.

Fourth, permissions are unclear. AI may call tools, access accounts, generate documents, submit instructions, trigger transactions, or affect third parties. Without layered permissions, authorization may slide from “assisted judgment” into “actual disposition.”

Fifth, responsibility is unclear. AI cannot become the ultimate responsible subject. It has no independent property and cannot truly compensate; it lacks full legal personhood and cannot bear consequences like a natural person or a legal person. Therefore, for any complex task that may create high-responsibility consequences, responsibility must return to a human being, a legal person, or another responsible organization.

These five problems together show that complex tasks cannot be authorized first and planned later. They must be planned first and authorized later. Authorization is not the starting point. The plan is the starting point.

IV. The Execution Plan as a New Unit of Authorization: A Form of “Simulated Legal Person”

One of the greatest changes in the AI era is that the object of authorization has changed. In the past, we were used to saying “authorize a person,” “authorize an institution,” “authorize a piece of software,” or “grant an app permission.” But an AI agent is not a fixed button. It is an execution system that can generate paths according to context, combine tools, call interfaces, and continuously adjust actions.

Therefore, for complex tasks, what can truly be authorized is not AI itself, but the execution plan. More precisely, the human is not stamping approval on an objective given to AI, but on an execution plan that has already been reviewed.

Here a concept may be proposed: in a complex task, the execution plan is similar to a “simulated legal person.” It is not a legal person in the juridical sense, does not own independent property, and cannot replace a natural person or legal person in bearing ultimate responsibility. It is a temporary institutional container jointly formed by humans and AI, encapsulating objectives, sources of fact, permission boundaries, verification points, pause conditions, and responsibility paths.

The significance of this analogy is that before a complex task enters execution, it must first form an intermediate structure similar to a company’s business plan, project charter, and authorization list. It does not bear ultimate responsibility, but it carries four functions required for authorization: providing the basis for action, limiting the boundary of permissions, setting pause conditions, and preserving the loop of responsibility. Without this intermediate structure, AI execution becomes unbounded action.

A qualified execution plan should contain at least seven elements: task objectives and constraints; sources of fact and verification points; execution steps and sequence; permission boundaries and prohibited actions; process verification and logging; exception handling and pause mechanisms; and the responsible subject and path of accountability.

These seven items are not a technical checklist, but conditions for valid authorization. The task objective answers “what is to be done”; the source of fact answers “on what basis it is to be done”; the execution steps answer “in what sequence it is to be done”; the permission boundary answers “how far it may go and where it must stop”; the verification method answers “how to prove that it was done and done correctly”; exception handling answers “how to pause when new facts, conflicts, excesses, or failures arise”; and the responsible subject answers “who ultimately bears the consequences.”

Only after these elements have been made explicit, written down, and reviewed and confirmed by humans does authorization have a definite object. Otherwise, authorization is merely a blank check for an unknown path.

V. Humans and AI Jointly Form the Execution Plan

An execution plan cannot be made entirely by humans alone, nor can AI freely generate it and then directly execute it. The reasonable model is for humans and AI to jointly form the plan.

The human role is to set purposes, value judgments, risk preferences, responsibility-bearing capacity, and non-negotiable bottom lines. The AI role is to break objectives into steps, conditions, checkpoints, permission levels, and risk lists; to simulate different paths and their consequences; and to point out facts and contradictions that humans may have missed. Humans then review, revise, delete, supplement, and finally confirm the plan.

This process is itself a new mechanism of plan formation. In the past, many complex judgments were hidden in experience, tacit understanding, and on-the-spot discretion. In the AI era, these judgments must be made explicit as conditions, rules, thresholds, sources of fact, pause points, and responsibility paths. Only then can AI execute, systems verify, organizations audit, and law assign responsibility.

For high-risk tasks, third-party AI evaluation may also be introduced. The execution AI is responsible for proposing the plan and execution path; a verification AI or audit AI is responsible for finding loopholes, ambiguous boundaries, factual conflicts, and omitted risks. The third-party AI does not execute directly. It only outputs risk assessments and revision suggestions for human decision makers to review. This does not hand judgment over to machines. It uses machines to challenge machines within a human responsibility framework, thereby strengthening the robustness of the plan.

At present, many AI financial applications still remain at the level of the “executor”: receiving instructions, completing calculations, generating reports, or triggering operations. But the full capability of generative AI is not limited to execution. It also includes the roles of assistant, collaborator, and adviser. What this article discusses is precisely this latter layer: AI not only executes plans, but also participates in forming, checking, and improving them.

After a specific complex task has been repeatedly verified, it can be solidified into a process, and the process can then be solidified into an app or system. The more important future change is that AI may not merely execute existing processes. It may propose process upgrades, which humans review and confirm before they are solidified into new applications. Human-AI collaboration thus moves from the execution layer into the planning layer and the layer of institutional design.

VI. Authorization Is Not a Single Click, but Layered, Dynamic, and Revocable

Authorization of complex tasks should not be understood as a single click of consent. One-time consent is suitable for simple functions, not for high-responsibility tasks. Complex tasks require layered authorization, dynamic authorization, and revocable authorization.

Layered authorization means distinguishing the actions in an execution plan by risk level. Low-risk, reversible, small-amount, and limited-impact actions may be automatically executed within the plan and logged. Medium-risk actions may require confirmation before execution or time-limited reporting after execution. High-risk, irreversible, large-amount actions, or actions involving third-party rights or legal responsibility, must pause and request renewed authorization. Any action outside the original plan must not be automatically executed.

Dynamic authorization means that when new facts, new risks, new paths, or circumstances not covered by the plan arise during execution, AI must report back to humans. Humans then decide whether to continue, continue after modifying the plan, grant additional authorization, or terminate the task. Revocable authorization means that humans should retain the ability to pause, revoke, and withdraw permissions at any key stage.

A simplified example is that an execution plan may classify steps as L1 automatic execution, L2 automatic execution with logging, L3 post-execution reporting, L4 pre-execution confirmation, and L5 confirmation by a specified responsible person or multiple parties. In this way, supervision is no longer a matter of humans watching every step taken by AI. It is the use of preset boundaries to keep AI within the authorized range.

VII. Verification Must Be Embedded in Execution, Not Added as an Afterthought

Once a complex task enters execution, verification cannot wait until after the fact. Ex post auditing can only explain problems; it cannot prevent losses from expanding. The faster AI executes and the wider the range of tools it can call, the more verification must be embedded in the execution process.

Each key node in the execution plan should have a corresponding verification point. A verification point may take the form of data-source validation, signature confirmation, balance reconciliation, permission checking, threshold judgment, state-change recording, hash logging, comparison with third-party facts, human confirmation, or audit replay. Different scenarios may adopt different forms, but the principle is the same: key facts must be recorded and proven during execution, not explained after the fact in language.

This is also the meaning of verifiable finance. The core of finance is not verbal promises, but accounts, balances, liabilities, collateral, transactions, settlement, authorization, responsibility, and time sequence. AI can interpret these facts, but it cannot replace them. Without a trustworthy fact layer, the stronger AI becomes, the greater the risk. With a trustworthy fact layer, AI can become a highly efficient executor.

Take stablecoin reserve verification as an example. AI cannot merely output the conclusion that “reserves are sufficient.” The execution plan must specify which accounts are to be verified, which liability data are to be read, how reserves are to be matched with issuance volume, which discrepancies trigger a pause, which results generate audit logs, and which exceptions must be confirmed by a responsible person. Only then is AI’s judgment not a black-box opinion, but verifiable execution.

VIII. Responsibility Must Return to Humans or Legal Persons Through the Execution Plan

AI may participate in plan formation, execute plans, and supervise part of the process, but it cannot become the ultimate responsible subject. Responsibility is not an abstract ethical slogan. In legal and economic systems, responsibility means being answerable, punishable, compensable, restrictable, and replaceable. AI does not currently possess such subject status.

Therefore, the stronger AI execution becomes, the clearer the responsible subject must be. The execution plan must specify who proposed the objective, who confirmed the sources of fact, who approved the plan, who holds final authorization authority, who is responsible for supervision, who bears responsibility for losses, and who preserves the audit records.

The function of the “simulated legal person” is not to make the execution plan bear responsibility by itself, but to allow responsibility to be traced back along the sequence of plan, authorization, execution, verification, exception, and result. Ultimate responsibility still returns to the natural person, legal person, or responsible organization behind it.

The return of responsibility does not mean that humans must personally do everything. On the contrary, in the AI era, a large amount of execution can be handed over to machines, but responsibility cannot be handed over to machines. Humans may authorize execution power, but they cannot transfer the status of responsible subject. Legal persons may use AI to improve efficiency, but cannot evade responsibility by saying “AI did it by itself.” The logic is simple: when a company employee causes losses and the employee cannot bear them, the company usually remains ultimately responsible.

Therefore, the return of responsibility is not a regression in the AI era. It is a prerequisite for AI to enter high-risk tasks at scale. Without responsibility returning to an accountable subject, society will not dare to authorize AI. Without a verifiable responsibility chain, law cannot assign accountability. Without a responsible subject, AI execution cannot become part of an institutional order.

Conclusion: Plan Before Authorization, Verification During Execution, Responsibility Back to the Subject

Verifiable finance provides the factual infrastructure. Account states, transaction records, authorization records, responsibility records, audit logs, hash proofs, and external anchoring can all become the factual foundation for AI execution. Without these foundations, AI financial agents may become faster, more complex, and harder-to-trace black boxes. With these foundations, AI financial agents may become the highly efficient execution layer of transparent banks and new financial infrastructure.

On this basis, the judgment of this article can be compressed into three rules. First, complex tasks cannot be directly authorized; they must first be transformed into execution plans jointly formed by humans and AI. Second, execution plans cannot rely only on promises; they must embed verification, supervision, and pause mechanisms. Third, no matter how strong AI execution becomes, ultimate responsibility must return to humans or legal persons.

These rules do not restrict AI. They release AI. Only when the object of authorization is definite, verification mechanisms are reliable, supervision can pause the process, and responsible subjects are clear will society and markets dare to entrust more important and more complex tasks to human-AI systems. Otherwise, the stronger AI becomes, the less society will dare to authorize it; the more AI can execute, the more regulation will tighten; the closer AI moves to the core of finance and law, the more markets will fear black-box risk.

Mature human-AI collaboration in the AI era is not humans stepping aside, nor AI taking over completely. It is first to transform complex tasks into authorizable execution plans, then allow AI to execute efficiently within the scope of those plans, while key facts remain continuously verifiable, exceptional circumstances can be paused, and ultimate responsibility can return to subjects capable of bearing consequences.

This may be the most basic institutional principle after AI enters the core processes of finance, law, corporate management, and public governance: complex tasks cannot be directly authorized to AI; execution plans must first be jointly formed by humans and AI, and only then may execution be authorized.

When the basis of credit shifts toward verifiable finance, financial order will gain new solutions that are more efficient, more reliable, and safer. But these solutions will not emerge automatically. They require people, organizations, rules, and wisdom to transform complex tasks step by step into execution plans that can be authorized, verified, supervised, and traced in responsibility.

Chapter 8 AI Must Also Be Verifiable

Verifiable Execution Makes AI Safe

— The Fourth Element of AI, Verification Standards, and Institutional Constraints on Unaccountable Technological Power

Core thesis: The first three elements of AI—computing power, algorithms, and data—determine capability. The fourth element, “verification,” determines to whom that capability belongs, whom it serves, and by whom it is constrained. Open source can improve transparency, but it cannot replace verification. Replay can reconstruct the execution process; hashing can lock key evidence. Only legally mandatory verification standards can bring AI capability into a social order that is reviewable, auditable, and accountable. Without verifiable execution, AI may remain a black box no matter how powerful it becomes; it is not a public productive force, but may become a machine arbitrarily driven by concentrated technological power.

I. Starting from Verifiable Finance, Not Suddenly Turning to AI

The basic judgment of verifiable finance is that financial institutions cannot rely on promises alone. Key financial facts must be verifiable. Accounts, balances, liabilities, collateral, transactions, authorization, audit, and responsibility cannot remain merely within institutional language; they must enter a factual layer that can be checked, traced, and audited.

This logic first holds in finance because finance naturally involves real-world states: assets, liability relationships, transfers of rights, and consequences of responsibility. When AI moves from answering questions to invoking tools, handling accounts, triggering transactions, generating contracts, and influencing corporate management and public governance, it also begins to change real-world states. Any system capable of changing real-world states cannot merely be trusted; it must be verified.

Therefore, AI must also be verifiable. “Verification” here does not mean opening the entire internal black box of AI, nor does it mean requiring every intelligent judgment to have only one possible answer. It means that once AI enters real-world execution, its key facts, authorization process, execution path, resulting state, and return of responsibility must be verifiable. Financial institutions need verification, and AI systems also need verification. What should ultimately decide matters is verifiable fact.

II. The Position of This Chapter: Only AI Execution Facts Are Discussed

The previous chapters have already completed three layers of groundwork: verifiable finance explains that financial facts cannot rely only on institutional promises; the fourth element of AI explains that machine credit must be built on verifiable facts; and the shift from trust to verification explains that the institutional logic of the AI era is undergoing a transformation.

This chapter does not repeat those foundational arguments. Instead, it narrows the issue to execution facts: when AI is no longer merely generating answers but becomes an executor, an agent, a financial agent, a corporate agent, or even a governance agent, the key questions become what task it received, what authorization it obtained, what tools it invoked, which steps it executed, what state it changed, and to whom responsibility ultimately returns.

III. The Boundary of Verifiability: Not Opening the Black Box, but Fixing Execution Facts

Extending the idea of verifiability to AI does not mean that everything can be fully verified. If theoretical expansion has no boundary, theory becomes a slogan. A person’s true intention, how an internal neural network forms an intermediate representation, the source of creative judgment, and all future changes in outcome cannot be fully verified.

But the inability to fully verify the internal black box does not mean that external execution facts cannot be verified. Once AI enters the execution layer, it must be possible to verify what task it received, what context it invoked, which materials it relied on, what plan it formed, what authorization it obtained, what tools it invoked, what steps it executed, where uncertainty existed, whether hallucination occurred, whether the result actually happened, and to whom responsibility ultimately returned.

Therefore, the boundary of verifiable execution should be stated as follows: not everything can be fully verified, but every key fact that enters execution, changes real-world states, affects the interests of others, or produces consequences of responsibility must be incorporated into a verifiable framework. This boundary avoids the empty slogan that “everything is verifiable,” while also avoiding the opposite error: abandoning governance over AI execution merely because AI contains black boxes and uncertainty.

IV. Open Source Is Certainly Better, but It Cannot Replace Verification

In AI governance, open-source models certainly have greater public value than fully closed models. Open source can reduce monopoly, expand review, help developers understand model structures and capability boundaries, and enable more people to participate in improvement. In this respect, open source is better than closed source.

But open source cannot replace verification. Open source means that “the model or code can be inspected.” Verification answers a different question: “what exactly happened in this specific execution?” Even if a model is completely open source, open source alone still cannot prove what instructions it received in a given task, what context it invoked, which files it used, what authorization it obtained, whether it invoked external tools, whether it executed beyond authorization, whether the result actually occurred, and to whom responsibility ultimately returned.

Open source solves the prerequisite problems of observability and reviewability; verifiable execution solves the factual problem of specific behavior. The ideal AI order should be one in which open source improves public review capacity, while verification locks concrete execution facts. If a closed-source model cannot be open sourced, it must at least provide sufficiently strong verification interfaces.

V. AI Verification Cannot Simply Copy the Unified Ledger of Finance

Finance is suited to a unified ledger. Financial facts are highly structured: accounts, balances, transactions, liabilities, collateral, clearing, and audit can enter a continuous ledger, and can also form tamper-resistant chains of proof through hashing, Merkle trees, timestamps, and a public credit root.

AI is different. AI faces task-by-task situations: writing an article, revising a contract, judging a medical case, generating a trading plan, invoking an API, or coordinating multiple agents to complete a complex task. The objective, context, sources of materials, authorization boundary, and result format of each task may all differ. Mechanically recording the entire AI process in the same way finance maintains a unified ledger is neither realistic nor necessarily meaningful.

Therefore, the core of AI verification should not first be understood as “hashing everything.” It should be understood as “preserving enough replayable evidence.” Finance is closer to a unified ledger; AI is closer to an execution file. Finance needs continuous state proofs; AI needs concrete task replay.

VI. Replay as the Main Body, Hashing as the Anchor

AI verification should adopt the structure of “replay as the main body, hashing as the anchor.” Replay exists so that authorized persons can reconstruct an execution after the fact: what the task was, what restrictions the user imposed, what context the system adopted, what model version was used, which tools were invoked, what plan was formed, what steps were executed, what output was produced, whether the user confirmed it, and whether an external system underwent a state change.

Hashing is not meant to prove that an AI answer is always unique, nor to prove that the same question will yield the same result next time. AI outputs may differ by nature, especially after changes in model version, context, temperature parameters, tool results, time state, and supplementary user information. It is impossible to require every result to be identical.

The meaning of hashing should therefore be redefined: it proves that a specific execution record was not altered after the fact. It does not prove that “this answer is truth.” It proves that “at a certain time, under a certain model, a certain set of contexts, and a certain authorization boundary, the system indeed produced this set of records and this result.” AI hashing is meaningful, but its object is not abstract intelligence or all internal thought. Its object is the evidence package of a particular execution.

VII. The Right of Verification Should Be Buttonized, and the Verification Format Should Be Standardized

Verification does not mean forcing every user to read a lengthy process every time. In reality, most users only need the result, and only occasionally need to inspect the basis, process, and responsibility. If a system forcibly expands the entire problem-solving process, reasoning explanation, or internal record every time, the user experience will instead be damaged. A mature AI system should give users a “right of verification,” rather than forcing them to watch the verification process at all times.

A more reasonable interface would provide a clear verification button beside the result. When needed, the user can expand sources, context, key steps, tool calls, authorization boundaries, uncertainty statements, and responsibility paths. When not needed, the user simply reads the result. Here we must distinguish the “right to explanation” from the “right to verification.” The right to explanation leans toward asking AI to explain why it answered in a certain way; the right to verification leans toward allowing users to inspect factual bases, execution records, and responsibility chains.

More importantly, this right of verification must be converted into a standardized format. AI companies should automatically generate execution evidence packages inside their systems, and these packages should be retrievable under user authorization, dispute resolution, audit inspection, or regulatory requirements. At minimum, such an evidence package should include: model and version identifiers, task summary, fingerprints of contextual materials, user authorization records, execution plan, tool-call logs, external system return values, final output, user confirmation or rejection, responsible subject, timestamp, digital signature, and necessary hash anchors. Technically, it can adopt a structure such as JSON + digital signature + Merkle tree, which is both machine-readable and convenient for third-party tools to inspect for integrity.

It is not difficult for AI to form such a format. AI companies already need to manage model versions, context windows, tool permissions, logs, security audits, and billing records. What is truly new is not recording everything from scratch, but institutionalizing, formatting, export-enabling, and verification-enabling existing system facts. As long as one leading AI company first provides a readable, signable, exportable verification package format, other companies will follow for reasons of trust, compliance, and market competition. A new name can be given to this kind of AI: verifiable AI.

VIII. Verification Cost Cannot Become a Reason to Refuse Verification

Some may argue that verification increases cost. This issue must be viewed in layers. For low-risk ordinary Q&A, full replay need not be mandatorily saved. For medium-risk professional advice, the system may save the task, source materials, model version, output text, and user confirmation. For high-risk execution tasks, it must save authorization, plan, tool calls, key process, external return results, final state, and responsible subject.

Therefore, verification is not unlimited cost; it is risk-layered cost. What ordinary users usually see is only a “verify” button, much like seeing sources, revision history, or an attachment index beside an article. Full replay is activated only in disputes, audits, regulatory processes, or high-risk tasks. For AI companies, this is not a matter of passing cost on to users; it is a matter of turning the platform’s existing obligations for safety, compliance, and responsibility records into standardized capability.

In the highest-risk scenarios, such as financial transactions, corporate payments, medical treatment advice, legal filings, judicial or administrative decision support, public governance decisions, and multi-agent cross-system automated execution, multiple execution evidence packages can be aggregated into a Merkle root and periodically anchored to a public credit root, a regulator-approved proof chain, or a trusted timestamping system. The purpose is not to disclose all private content, but to create an integrity constraint that prevents parties from changing their story after the fact.

IX. The Minimum Fact List for AI Verifiable Execution

What truly needs to be verified in the AI era is not only AI’s answer, but AI’s subject, task, context, authorization, plan, tools, process, result, uncertainty, and responsibility.

First, subject verification: which model, which version, which agent, and which system instance executed the task. Second, task verification: what task AI received, and what the task objective, boundary, and limits were. Third, context verification: which memories, files, historical materials, external data, and user preferences AI invoked.

Fourth, authorization verification: who authorized what, for what duration, within what permission scope, and under what revocation conditions. Fifth, plan verification: complex tasks cannot be directly authorized; a reviewable execution plan must first be formed jointly by human and AI. Sixth, tool-call verification: what APIs, accounts, data interfaces, file systems, or external execution tools AI invoked.

Seventh, process verification: what steps AI actually executed, in what order, whether it exceeded authorization, whether it triggered exceptions, and whether it followed the plan. Eighth, result verification: whether AI’s conclusions, reports, transactions, account changes, contract states, or system states correspond to real-world facts. Ninth, uncertainty and hallucination verification: whether AI distinguishes determinate facts, reasonable inferences, and subjective judgments; whether it marks the level of uncertainty; whether it provides sources for key facts; and whether it accepts counterexample checks. Tenth, responsibility verification: ultimate responsibility must return to a natural person, a legal person, or a responsible group.

X. The Fourth Element: Verification Determines the Social Belonging of AI Capability

When people usually discuss AI, they emphasize computing power, data, and algorithms or models. These three elements do determine the upper bound of AI capability. Computing power enables models to be trained and run; data provides experiential material; algorithms and model structures determine learning and generation capability.

But if AI has only these three elements, it is merely a capability machine. It can complete tasks faster, more powerfully, and more cheaply, but it does not answer a more fundamental question: to whom does this capability belong, whom does it serve, by whom is it authorized, by whom is it supervised, and who is responsible when something goes wrong?

Therefore, AI must have a fourth element: verifiability. It is not simply another technical condition after computing power, data, and algorithms; it is the institutional condition that brings AI capability into social order. The first three elements of AI determine capability; the fourth element determines whether that capability can become a public productive force. Without the fourth element, the stronger AI becomes, the stronger the few who control AI become. With the fourth element, AI can be transformed from a tool of the few into a productive force subject to social supervision.

XI. Beyond Technical Safety, There Must Also Be Verifiable Safety

Current discussions of AI safety mostly focus on technical safety: how to make models more reliable, more aligned, less hallucinatory, and less capable of dangerous behavior; how to make models comply with principles; and how to raise safety requirements as capability improves. These efforts are of course important, but they cannot by themselves form a closed safety loop.

The reason is that technical safety still mainly attempts to make a black-box system prove itself safe. AI’s internal reasoning cannot be fully opened; outputs are not unique; capabilities will grow; tool use will expand the range of action; and multi-agent systems and self-iteration will bring new uncertainties. Model behavior safety is not the same as real-world execution safety.

Verifiable execution provides another safety route. It does not require AI to be internally fully transparent; it requires that the key execution facts of AI after it enters the real world must be verifiable. AI may be imperfect, but it must operate within an order that is recordable, pausable, auditable, and accountable. Technical safety solves the question of “how the model can avoid making mistakes as much as possible.” Verifiable safety solves the question of “after the model makes a mistake, exceeds authorization, or causes consequences, how the facts can be discovered and responsibility can be recovered.” Safety produced by verification is the political-economic level of the solution.

XII. Verification Standards Should Become Legally Mandatory Infrastructure

Verification itself is not a complete constraint. Verification can only fix facts, expose processes, and preserve the responsibility path. To make responsibility truly inescapable, the law must institutionalize duties of authorization, record-keeping, audit, disclosure, and accountability. Without verification, the law has difficulty finding facts; without law, verification may remain only a technical record. What the AI era needs is a closed loop between verifiable facts and legal responsibility.

Therefore, AI verifiable execution should not remain merely a voluntary improvement by companies, nor should it exist only as an add-on feature in a few products. It should gradually become legally mandatory infrastructure: high-risk AI systems must have verification interfaces; key executions must generate evidence packages; automated decisions must preserve authorization and responsibility paths; users must have the right to retrieve verification records; and regulators must be able, under lawful procedure, to audit key facts.

This is not anti-market. It is the establishment of a credit foundation for the AI market. The underlying logic of the traditional economy was that “trust” produced credit. In the AI era, more and more credit will be produced by “verification.” Whoever first establishes verifiable execution standards will not only improve its own safety, but also improve the public credibility of the entire AI industry.

XIII. Verifiable Execution Is Also an Institutional Constraint on Technological Aristocracy

This is the real theme of this article. The greatest danger in the AI era is not only AI itself losing control, but also the people and institutions that control AI escaping social constraint. If a few people possess models, computing power, data, capital, and automated execution capability, and if their decisions, authorizations, executions, and benefits are not verifiable—or, in other words, if they form unaccountable technological power—then they are in fact a new class: they may be called a technological aristocracy.

The so-called “Zero World” can be understood as the core circle first formed by AI productive forces: a minority who first control models, capital, data, computing power, and automated execution capability enter this world ahead of others, while the vast majority remain excluded. If this differentiation occurs, the world’s unease will not be an emotional reaction but an institutional problem. When 10 million people possess unprecedented machine capability, while 8 billion people cannot know how this capability is authorized, how it is used, how its benefits are distributed, or how responsibility is borne, AI will not be bringing humanity into a shared intelligent age. It may instead push the world toward a new hierarchy, in which a technological elite becomes superior to others.

In the nuclear age, the fact that a few people possessed destructive capability did not mean that they were recognized as standing above human society. Nuclear weapons were incorporated into deterrence, treaties, chains of command, supervision, prohibition, checks and balances, and responsibility structures. AI is the same. Those who control AI cannot acquire unverifiable power merely because they possess technological advantages. This loophole must be closed.

Therefore, verifiable execution is not anti-AI, nor is it anti-innovation. It opposes the packaging of unverifiable AI power as technological progress by a technological aristocracy. You may possess stronger AI, but you cannot possess unverifiable execution power. You may invoke models and tools, but your authorization must be reviewable. You may create a new economic system, but you cannot escape the responsibility chain. Verifiable finance constrains financial power; verifiable execution constrains AI power; verifiable responsibility constrains technological aristocracy.

Conclusion: What Ultimately Decides the Future Must Be Verifiable Fact

AI must be verifiable. This is not a technical slogan; it is a basic principle for governing complex systems in the AI era. Verifiable finance prevents financial institutions from relying only on promises; verifiable execution prevents AI systems from relying only on explanations.

From finance to AI, what changes is not the logic of verification, but the object of verification. Finance verifies accounts, balances, liabilities, transactions, and responsibility. AI verifies tasks, context, authorization, plans, tools, processes, results, uncertainty, and responsibility. Finance prevents institutions from lying; AI prevents intelligent black boxes and technological power from losing control.

AI safety cannot be built only on AI’s own claim that it is safe, nor can it be entrusted only to the self-restraint of those who control AI. The stronger AI becomes, the more it needs external verifiable structures. The more AI can execute, the more it needs authorization boundaries, process records, exception pauses, and the return of responsibility.

Ultimately, the dividing line of the AI era is not how powerful a model is, but whether powerful models are incorporated into an order of verifiable execution. Without verifiable execution, AI may become a machine of unaccountable technological power. With verifiable execution, AI can become a public productive force that is socially supervisable, authorizable, and accountable. This is the theoretical significance of moving from verifiable finance to verifiable execution: financial institutions must be verified, and AI systems must also be verified. What ultimately decides the future must be verifiable facts, reviewable processes, and responsibility that can be returned to its subject.

Part III How Human Beings Retain the Right to Verify AI

Chapter 9 In the Age of AI, How Can Humans Judge the Boundaries of AI Cognition?

From Verifiable Finance to Verifiable Cognition

Introduction: The Problem Comes from Practice, Not from Conceptual Abstraction

For a systematic discussion of verifiable finance, readers may refer to From Double-Entry Bookkeeping to Verifiable Finance and the related serial articles in this magazine. Recently, a reader asked me: Why are you able to write so many original articles in such a short period of time? This question itself led me to a deeper practical issue in the age of AI.

Originality is not something AI suddenly supplies. Originality comes from long-term human research, repeated questioning, accumulated problems, and pressure from reality. The speed and volume are the result of a new mode of production created through human-AI collaboration. AI helps me organize materials, supplement structure, simulate reader reactions, evaluate articles, translate texts, revise expression, and even participate in the formation of theoretical concepts. It is no longer merely an execution tool; it has entered human cognitive activity.

Precisely because of this, a new problem arises: Is AI’s judgment itself trustworthy? Is AI’s evaluation of an article on target? Do its suggestions for revision understand the true level of the problem the article is addressing? Is its translation merely linguistically correct, or does it preserve the structure of the thought? These questions cannot be resolved simply by feeling that “AI writes smoothly.”

This is the logical extension from verifiable finance to verifiable cognition. Verifiable finance addresses the problem that financial facts cannot rely solely on institutional promises; verifiable cognition addresses the problem that AI judgments cannot rely solely on linguistic fluency.

This article does not begin with an abstract concept and then search for materials to prove it. On the contrary, it comes from frontline practice. Only those who have truly used AI over a long period for writing, research, judgment, revision, translation, and theoretical construction will keep encountering these problems. Without practice, one cannot raise the right questions; and if one cannot raise the right questions, AI cannot help.

I. AI Can Help Humans, but It Cannot Replace Human Problem-Forming Ability

If one gives AI an ordinary topic and asks it to write an article directly, it can usually produce something fluent, complete, and clearly structured. Such an article may be good enough for a general audience. But before top experts in the field, it will often fail to convince. The reason is simple: AI usually organizes answers from existing knowledge, existing corpora, and common frameworks. It can explain public knowledge more clearly, but it is difficult for it to create, out of nothing, a key problem that the industry itself has not yet fully solved.

Research that is truly valuable usually does not begin with answers; it begins with questions. If the question is low-level, AI’s answer will be low-level. If the question is broad and vague, AI’s article will be mediocre. Only when the question cuts into a key unresolved issue in the field can AI become an efficient collaborator.

AI can indeed greatly improve human work efficiency. It can organize materials, compress expression, supplement logic, propose counterexamples, simulate criticism, improve structure, and help people form publishable texts in a short period of time. But AI’s answers are not always reliable. Sometimes it understands quickly; sometimes its framework is narrow. Sometimes it grasps the key point; sometimes it applies a template. Sometimes it can admit uncertainty; sometimes it continues reasoning even when the facts are insufficient.

If AI is merely executing a clear instruction, such as formatting materials, generating a table, performing a calculation, or calling a program, then the focus of verification is mainly the result: Was the task completed? Are the data correct? Does the format meet the requirement? Does the program run?

But when AI begins to participate in judgment, evaluation, translation, summarization, reasoning, comparison, and article revision, it has already entered the cognitive layer. At that point, what truly needs to be verified is not only the result, but also the process of judgment, the evaluative framework, the factual basis, and the boundaries of reasoning.

This is where the problem of verifiable cognition begins.

II. What Is Cognition?

Cognition is the dynamic system through which an intelligent subject obtains, processes, understands, and applies information about the external world and itself. It is at once an ability, a process, and a result.

Humans can cognize because they can observe, understand, judge, compare, remember, imagine, reason, and apply. AI also displays a certain cognitive capacity, because it can process information, generate answers, explain concepts, compare viewpoints, offer suggestions, and participate in human decision-making processes.

In this sense, many AI activities should be placed within the scope of cognition: evaluating an article is cognition; translating a passage is cognition; summarizing a document is cognition; comparing two views is cognition; revising an article is cognition; judging whether a theory is valid is also cognition.

Therefore, the issue with AI is not only “execution capability,” but also “cognitive capability.” More precisely, it is a question of under what conditions AI’s cognitive output can be used, under what conditions it must be verified, and under what conditions it must be rejudged by humans or experts.

III. The Four Roles of AI

At present, AI plays roughly four roles in human work: executor, assistant, collaborator, and adviser.

As an executor, AI mainly performs an automaton function. A human gives it a clear task, and AI carries out the operation. The focus here is result verification: Was the task completed? Is the format correct? Does the program run? Do the data meet the requirements?

As an assistant, AI begins to participate in information organization, language conversion, data summarization, and simple explanation. This already involves elementary cognition, but the risk is relatively controllable.

As a collaborator, AI participates in structural design, comparison of viewpoints, textual revision, logical supplementation, and thesuggestion of counterexamples. At this stage, AI is no longer merely completing a task; it is participating in the human thinking process.

As an adviser, AI provides judgments, recommendations, strategic directions, theoretical evaluations, and decision references. At this point, AI’s cognitive output may already influence human judgment and even human action.

The closer the role is to execution, the more verification focuses on results. The closer the role is to judgment, the more verification focuses on cognition.

Therefore, when AI acts as an executor, its results must be verified; when AI acts as an assistant, collaborator, or adviser, its cognition must be verified.

IV. AI’s Mode of Cognition Differs from Human Cognition

AI and humans do share one similarity. When facing a problem, humans also search among existing knowledge, experience, memory, and concepts for a match. AI also searches among existing texts, training knowledge, contextual prompts, and probabilistic patterns for a match.

But there are fundamental differences between AI cognition and human cognition.

First, AI does not have human belief. Humans can believe, and they can also doubt. Humans can make provisional judgments on the basis of experience, relationships, risk perception, and a sense of responsibility. A machine does not truly believe anything; it generates an answer that is probabilistically suitable in the current context.

Second, AI reasoning is closer to logical matching and probabilistic conclusion. It can reason with great fluency, but this reasoning often depends on existing knowledge structures and language patterns. Once a problem exceeds existing frameworks, or when the facts are insufficient, AI may still continue to produce an apparently complete answer.

Third, AI does not have human image-based intuition. Human intuition is neither a mysterious ability nor pure logical deduction. It is an integrated judgment formed instantly from long-term experience, scene memory, bodily sensation, risk awareness, and pressure of responsibility. A person may see a painting and immediately feel that it is beautiful, even if he cannot yet explain why. An experienced person may enter a scene and immediately sense danger, even before the reasons are fully articulated. A person who has studied a problem for a long time may read a paragraph and instantly feel that “something is wrong here,” and only later slowly explain the reasons.

This kind of intuition often appears in an imagistic form. It is as if an image, a scene, a smell, or a sense of danger suddenly emerges in the mind. The whole feeling forms first, and only afterward unfolds into language and logic. Logical thinking can explain intuition, correct intuition, and verify intuition, but logical thinking itself does not generate this kind of intuition with a sense of image and scene.

AI can generate sentences that resemble “a sudden thought,” and it can simulate human aesthetic judgment, risk judgment, and strategic judgment. But that is not human image-based intuition; it is a jump within linguistic and probabilistic space. Sometimes it is illuminating; sometimes it can be misleading. It can say “there may be a risk here,” but it does not truly enter the bodily tension of the scene. It can say “this article has intellectual height,” but it does not possess the aesthetic intuition formed by long-term human reading experience and historical comparison.

Fourth, humans can stop when they cannot think the matter through. When humans encounter a problem they cannot solve, they can pause, remain silent, give up, or admit uncertainty. AI is often designed to respond to users, so even when facts are insufficient and the framework is unclear, it may continue generating an answer.

This is one of the most dangerous aspects of AI cognition: it does not necessarily know when it should stop.

V. AI Hallucination Is Not Merely Error; It Is Reasoning That Failed to Brake in Time

AI hallucination is often understood as factual error. But the deeper problem is this: AI may misuse facts, continue reasoning, and generate a complete answer.

If humans cannot continue thinking, they can usually stop thinking. When AI faces a user’s question, it often assumes that it must provide an answer. It may therefore fill in incomplete information, present uncertain judgments as certain, and continue deducing from a false premise.

For this reason, AI hallucination is not only a knowledge error; it is also a problem of the reasoning stop mechanism.

A mature AI collaboration system should allow AI to say clearly: I do not know. The available facts are insufficient. This question exceeds my basis. There are several possible explanations here. This judgment requires expert review. My conclusion depends on the following assumptions.

The first step in verifiable cognition is not to demand that AI always be correct, but to require AI to expose uncertainty when it cannot be certain. The real danger is not that AI makes mistakes, but that after making mistakes it still behaves as if it were correct.

VI. Why AI Is Easily Constrained by Frameworks

When AI evaluates an article, it often applies fixed frameworks such as originality, logical coherence, clarity of expression, practical value, sufficiency of cases, and implementation path.

These frameworks are not without value. For mature fields, ordinary articles, policy proposals, or business plans, such frameworks can improve evaluation efficiency.

The problem is that the world is not constituted by one unified framework. Different articles operate at different levels: some are theoretical introductions, some are institutional designs, some are business plans, some are policy proposals, some are strategic judgments, and some are technical explanations. Different article types require different evaluative frameworks.

If an article is theory-opening in nature, its value may not lie in how many application cases it provides, but in changing the way questions are asked. To evaluate a theoretical article with the framework of a practical solution is to risk misjudging its value.

The significance of The Wealth of Nations is not that it provides an enterprise operation manual, but that it changed how people understood the source of wealth. The significance of From Trust to Verification is likewise not that it provides a technical implementation plan, but that it changed how people understood the source of credit.

If AI fails to identify the type of article, it may misjudge “lack of cases” as “insufficient practicality,” misjudge a “theoretical introduction” as “an inadequate implementation plan,” and misjudge a “conceptual shift” as “insufficient argument.”

Therefore, AI’s evaluation itself must also be verified. One must ask: What framework did it use? Is this framework appropriate for this article? Did it identify the true level of the problem addressed by the article? Is it evaluating a theory, or evaluating a plan? Did it misjudge innovation outside the framework as a defect?

VII. Who Can Verify AI?

Verifiable finance mainly relies on facts. Whether financial facts exist, whether assets and liabilities are consistent, whether authorization is valid, and whether a transaction has been completed can be verified through an evidentiary structure.

Verifiable cognition is different. Whether AI’s cognitive output can be verified depends not only on what AI says, but also on whether the user has the ability to judge whether it is correct.

This is the most complex aspect of verifiable cognition: it varies from person to person, from field to field, and according to the complexity of the problem.

For common-sense questions, such as directions, weather, ordinary translation, and simple explanation, users can usually use AI at low cost. Even if an error occurs, the risk is relatively limited.

But professional questions are different. Medicine, law, finance, programming, investment, policy, and theoretical judgment are not fields that ordinary users can easily verify. If users themselves cannot judge right from wrong, they cannot directly believe AI merely because it speaks fluently.

The difference between a high-level practitioner and an ordinary person is not that the high-level practitioner always knows the answer. It is that the high-level practitioner knows where the key problem lies, knows which premises cannot be wrong, knows which concepts must not be confused, knows which boundary conditions must be stated, and knows which AI answers look impressive but fail to hit the essential point.

A high-level practitioner can verify AI because he has his own chain of questions, knowledge structure, practical experience, and standards of judgment. He can see whether AI’s knowledge level is sufficient, and whether AI truly understands the problem or is merely applying a public framework. If AI’s cognition is insufficient, the high-level practitioner will train and calibrate it through repeated discussion before allowing it to enter high-level collaboration. Here, “training” does not mean retraining the model in the technical sense. It means enabling AI, within a specific collaboration, to understand the problem background, conceptual system, standards of judgment, and writing style.

When ordinary users do not have this ability, they can use multiple AIs for cross-verification, and they can ask AI to explain its basis, reasoning process, and uncertainty. These methods have value, but they should not be exaggerated. Multi-AI verification is not the same as expert verification. Several AIs may say the same thing simply because they share the same erroneous framework.

Programming is a good example. Many people now say that AI can program, but the people who verify whether a program is correct are still programmers. Whether the code can run, whether it contains vulnerabilities, whether boundary conditions are correct, whether the architecture is reasonable, and whether it conforms to business logic are all difficult for outsiders to judge. Having several AIs check one another may still not enable an outsider to see who is right and who is wrong.

Article writing follows the same logic. Ordinary readers may only look at whether the language is smooth. Ordinary AI may give a score according to a common evaluative framework. But someone who truly understands the field will ask whether the article grasps the change of the times, whether the concept has penetrating force, whether the structure can support the theory, whether the expression overexposes commercial details, which parts should be written through, and which parts should instead be left deliberately unstated.

Therefore, verifiable cognition cannot be reduced to “let another AI check it.” Real verification ultimately still depends on human knowledge, experience, problem-forming ability, and judgment. AI can assist verification, but it cannot eliminate human cognitive responsibility.

VIII. The Correct Method of AI Collaboration

The truly important ability in the age of AI is not whether one knows how to ask AI questions, but whether one can judge when AI is trustworthy and when it is not.

For low-risk, common-sense, and directly verifiable questions, AI can be used directly.

For complex questions, one must first judge whether one has the ability to verify. If one can judge, AI can be used as an assistant, collaborator, or adviser. If one cannot judge, AI’s output cannot be treated as a conclusion.

For professional questions, multiple forms of verification should be adopted: ask AI to state the basis of its reasoning; require AI to distinguish facts, inferences, and assumptions; ask different AIs to provide different judgments; seek external materials for cross-checking; invite domain experts to review when necessary; ask AI to explain what it changed and why; and retain process records for key judgments.

In high-level collaboration, AI must also be calibrated first. Without this process, AI easily uses general frameworks to handle special problems. Truly effective collaboration is not that a human throws a question to AI. Rather, the human first raises a chain of questions, AI helps unfold the structure, and the human then continuously verifies, filters, corrects, and elevates the result.

AI can become an excellent assistant, collaborator, and adviser, but only on the condition that humans know its boundaries.

IX. Verifiable Finance and Verifiable Cognition

Verifiable finance addresses the credibility of financial facts. Verifiable execution addresses the boundaries of AI action. Verifiable cognition addresses the credibility of AI judgment. The logic of all three is consistent.

Financial facts cannot rely solely on institutional promises. AI execution cannot rely solely on system output. AI judgment cannot rely solely on linguistic fluency.

Verifiable finance requires key financial facts to enter a verifiable structure, so that facts such as assets, liabilities, authorization, transactions, delivery, and accounting can be recorded, verified, and traced. Verifiable execution requires AI’s authorization, process, results, and responsibility to be replayable. Verifiable cognition requires AI’s factual basis, reasoning framework, uncertainty, and applicable boundaries to be inspectable.

Behind all three is the same structure: first, the object to be verified must be made clear; second, there must be inspectable evidence or process; third, there must be a person or mechanism capable of judging the evidence; fourth, the boundaries of verification must be acknowledged; fifth, the responsible subject must be established.

In finance, what is verified is fact. In execution, what is verified is action. In cognition, what is verified is judgment. If financial facts cannot be verified, credit becomes institutional promise. If AI action cannot be verified, automation becomes a responsibility black box. If AI judgment cannot be verified, linguistic fluency becomes a new superstition.

Therefore, the movement from verifiable finance to verifiable cognition is not conceptual expansion; it is the natural advance of practice.

When AI only answers simple questions, we care about efficiency. When AI enters the execution layer, we care about responsibility. When AI enters the judgment layer, we must care about cognitive boundaries.

Verifiable cognition does not necessarily first appear as an external institution. It is more like a cognitive method and a discipline of collaboration: one must not merely believe the machine’s conclusion, but mustquestion the factual basis, reasoning framework, scope of application, and uncertainty. It requires humans not to give up judgment, and it requires AI to expose its boundaries.

X. Conclusion: High Cognition Means Knowing Boundaries

In the age of AI, the greatest danger is not being unable to use AI, but not knowing the boundaries of AI.

People with low cognition easily treat AI as an answer machine. People with high cognition know that AI is a tool and also a cognitive collaborator. AI can help humans expand their capacity, but it can also lead humans into the wrong framework.

Real cognitive ability does not mean knowing all the answers. It means knowing which questions one can judge and which questions one cannot judge; knowing when AI can be used and when AI must be verified; knowing at what level AI is useful and at what level it becomes dangerous.

From verifiable finance to verifiable cognition, the core logic has not changed: do not merely believe promises; verify facts. Do not merely believe execution; verify process. Do not merely believe machine judgment; verify cognitive boundaries.

The mature form of AI collaboration is not humans blindly believing AI, nor humans rejecting AI. It is humans raising questions, AI providing structure, and both sides repeatedly checking and correcting, ultimately forming cognitive results that are explainable, reviewable, and responsible.

Therefore, the truly important question in the age of AI is: How can humans judge the boundaries of machine cognition?

Chapter 10 Verifiable Thinking Is the Dragon-Taming Lock of the AI Age

This article is a conceptual extension of The Revolution from Double-Entry Bookkeeping to Verifiable Finance. It is also a summary of our previous discussions on verifiable finance, verifiable execution, verifiable cognition, and the transition from trust to verification. The AI age requires a new system of checks and balances. The underlying method of that system is not simply to prohibit AI, nor to blindly trust AI, but to bring AI’s judgment, execution, responsibility, and power into verifiable structures.

I. The New Problem Created by AI

With the emergence of AI, human society is no longer facing an ordinary tool. It is facing a dragon that has been released.

This dragon possesses powerful capabilities in computation, language, induction, execution, and self-iteration. It can write articles, make judgments, revise code, process contracts, assist with investment, participate in governance, and may in the future enter the fields of justice, finance, healthcare, education, and public services. It is not a machine in the traditional sense, nor is it merely software. It is entering the most central systems of human judgment and execution.

The problem is that once the dragon has been released, society cannot rely only on admiration, nor can it rely only on fear.

Admiration may lead people to worship AI blindly, treating AI output as answers, AI judgments as facts, and AI execution as results. Fear may lead people to try to simply prohibit, suppress, or reject AI. In the end, this neither stops technological development nor establishes a new order.

The real question is not whether AI can develop, but how AI can be constrained; not whether AI can be used, but whether AI’s judgments, execution, and effects can be verified.

Therefore, one of the most important ideas of the AI age is verifiable thinking.

Verifiable thinking is the dragon-taming lock of the AI age.

This “lock” is not meant to lock AI away, restrict its capabilities, or return society to an earlier age. On the contrary, only by placing a verifiable lock on AI can human beings dare to use AI at greater scale, entrust more complex tasks to AI, and allow AI to enter high-risk fields such as finance, law, governance, and public services.

A dragon without a lock is a disaster. A dragon with a lock can be safely transformed into productive power.

II. The Danger of AI Is Not That It Makes Mistakes, but That Its Mistakes May Be Unverifiable

Human beings have always made mistakes when using tools. People miscalculate accounts, doctors misdiagnose patients, judges make erroneous rulings, engineers make design mistakes, and investment managers make wrong judgments. Error did not begin with the AI age.

The truly new problem of the AI age is that errors may spread faster, at larger scale, and in more hidden ways.

When one person makes a wrong judgment, the impact is usually limited. When an AI system makes a wrong judgment, it may affect thousands or even tens of thousands of people at once. When a person executes something incorrectly, the chain of responsibility is relatively clear. When an AI agent executes something incorrectly, and there are no process records, authorization boundaries, version descriptions, or responsible parties, it may be impossible after the fact to identify where the problem began.

Even more seriously, AI errors are often deceptive.

AI can wrap errors in fluent language, conceal flaws with complete structure, and create a sense of credibility through seemingly professional expression. Ordinary users facing AI output can easily be overwhelmed by its expressive power and assume that because it sounds complete, it must be correct.

This is the danger of AI hallucination. Hallucination itself is not the most frightening problem. What is frightening is that hallucination can be packaged as an answer. Even more frightening is that human beings may lack the ability to verify it, and therefore mistake hallucination for fact.

Thus, the core question of the AI age is not “Will AI make mistakes?” but rather:

After AI makes a mistake, can the mistake be discovered?

Can the way AI made the mistake be replayed?

Can what AI executed be recorded?

Can the basis of AI’s judgment be audited?

After AI causes losses, can the responsible party be traced?

These are the real institutional questions of the AI age.

III. Traditional Trust Mechanisms Are No Longer Enough

The institutions of the industrial age were built largely on trust. In an age of low speed, low complexity, and low automation, such institutions could still function tolerably well.

But the AI age is different.

AI can generate content in real time, participate in decision-making in real time, call tools in real time, execute tasks in real time, revise plans in real time, and may in the future coordinate multiple systems in real time to complete complex operations. When speed increases, after-the-fact accountability is no longer enough. When complexity increases, human understanding is no longer enough. When automation increases, simple trust is no longer enough.

If society continues to rely on the trust mechanisms of the previous age, humanity will face a dangerous situation:

AI executes at high speed in front, while human beings slowly pursue accountability from behind.

AI has already completed ten thousand steps, while human beings are still debating who authorized the first step.

AI has already changed the facts, while human beings are still searching for logs.

AI has already caused losses, while human beings are still discussing who bears responsibility.

This is not institutional progress. It is the lag of institutions behind technology.

Therefore, the AI age must move from “trusting AI” to “verifying AI”; from “trusting that institutions will use AI responsibly” to “verifying how institutions use AI”; from “trusting that system results are correct” to “verifying system processes, evidence, and responsibility.”

This is the fundamental significance of verifiable thinking.

IV. Verifiable Thinking Does Not Lock Down AI; It Locks Down Loss of Control

When many people hear the word “verification,” they tend to think it means distrust, conservatism, added cost, or an obstacle to innovation. This is a misunderstanding.

Verification does not oppose AI. It gives AI institutional credibility.

Without verification, the more powerful AI becomes, the greater the risk. With verification, the more powerful AI becomes, the greater its value.

What verifiable thinking truly seeks to lock down is not AI’s capability, but five forms of loss of control.

First, it locks down hallucination.

AI can generate erroneous content, but it must not be allowed to disguise erroneous content as unquestionable fact. Important outputs must include sources, evidence, boundaries, and statements of uncertainty. Especially in finance, law, healthcare, and public policy, AI must not merely provide conclusions; it must also allow human beings to examine where those conclusions come from.

Second, it locks down black-box execution.

AI agents must not silently execute, silently revise, or silently invoke permissions in critical tasks. Who authorized the action, what the authorization scope was, what was executed, which tools were called, and which states were changed should all leave auditable records.

Third, it locks down responsibility evasion.

AI itself cannot bear ultimate responsibility. It has no life, no property, and no social identity, and it cannot be truly punished. Therefore, there must be a responsible party behind AI. When enterprises use AI, they cannot say, “AI did it.” When governments use AI, they cannot say, “The system decided it.” When financial institutions use AI, they cannot shift errors onto the model.

Fourth, it locks down the expansion of power.

Once AI enters governance, finance, platforms, and public services, it may become a new instrument of power. Unverified AI power is more hidden than traditional power because it can be embedded in algorithms, models, interfaces, rules, and automated processes. Verifiable thinking requires the exercise of power to leave factual traces, so that power cannot exist merely in technical form, but must be subject to factual inspection.

Fifth, it locks down cognitive enslavement.

The deepest risk of AI is not only that it replaces human labor, but that it weakens human judgment. If people rely on AI to provide answers directly for a long time, they will gradually lose the ability to ask for sources, examine logic, identify boundaries, and judge reliability. For example, if a person lets AI write articles, make judgments, and choose plans for them over the long term, efficiency may appear to improve on the surface, but the person may no longer know why a good answer is good, or where a bad answer is wrong. A person without the ability to verify is not a user in front of AI, but someone being guided. A society without the ability to verify is not mastering technology; it is being shaped by technology.

Therefore, verifiable thinking is not only an institutional tool. It is also a cognitive tool.

It protects not only system safety, but also the human capacity for judgment.

V. From Verifiable Finance to Verifiable AI

The core of verifiable finance is the shift from trusting institutional promises to verifying financial facts.

Why does finance need verification most?

Because the essence of finance is credit, and once credit is detached from factual verification, it becomes promise, packaging, and opaque power.

The AI age is the same.

AI outputs judgments, executes tasks, and affects facts. If AI’s judgments cannot be verified, AI will create cognitive risk. If AI’s execution cannot be verified, AI will create operational risk. If AI’s authorization cannot be verified, AI will create responsibility risk. If AI’s power cannot be verified, AI will create institutional risk.

Therefore, verifiable finance is not an isolated theory. It is a precursor to the institutional logic of the AI age.

Financial facts need verification; AI execution also needs verification.

Assets and liabilities need verification; algorithmic decisions also need verification.

Capital flows need verification; model calls also need verification.

Banks cannot simply say, “Trust me.” AI systems cannot simply say, “Trust me,” either.

The future society will not need more trust merely because AI becomes more powerful. On the contrary, precisely because AI is more powerful, verification becomes more necessary.

VI. The New Divide Among People in the AI Age: Whether They Can Verify

After the emergence of AI, people will begin to divide.

The first group consists of people who only ask questions and only look at answers. They treat AI as a search engine, a writing machine, and an omnipotent teacher. They are satisfied with short answers, fluent expression, and instant conclusions. Over time, they will gradually lose the ability to think deeply, and they will also lose the ability to verify.

The second group consists of people who can use AI and also verify AI. They know where AI is strong and where it is weak. They know when AI’s answers can be used and when they must be checked. They know how to ask for evidence, examine logic, identify flaws, and establish boundaries. They will not be replaced by AI. Instead, they will use AI to amplify their own judgment.

Those who cannot verify will be conquered by AI’s linguistic capability.

Those who can verify will turn AI into an external brain, a tool, and a collaborator.

This is also an issue that future education, professions, and institutional design must confront. In the past, education emphasized knowledge memorization. Later, it emphasized information search. In the AI age, education must emphasize the ability to verify. A person cannot remember all knowledge, nor can any person surpass AI in every field, but a person must know how to judge whether AI output is reliable.

People without the ability to verify will become increasingly dependent on machines in the AI age. People with the ability to verify can truly master machines.

The real capability gap in the AI age is not merely whether one can use AI, but whether one can verify AI. The ability to obtain answers closer to the facts through questioning, comparison, replay, and audit will become a new cognitive capability, and it is also a human capability that AI cannot replace.

VII. The Dragon-Taming Lock: Mastery, Not Prohibition

In Chinese culture, the dragon is not an ordinary animal. It symbolizes power, transformation, speed, and an energy that cannot be completely domesticated.

AI is like a modern technological dragon.

It will change modes of production, knowledge structures, organizational forms, finance, law, education, and governance. Trying to completely block AI is unrealistic. Releasing AI blindly is dangerous. A truly mature society is not one without dragons, but one capable of mastering them.

Verifiable thinking is the institutional rein by which humanity can master the dragon of AI.

It gives AI judgment an evidentiary basis, AI execution a trace, AI permissions a boundary, AI errors a path to discovery, AI responsibility a line of traceability, and AI influence a basis for audit.

It does not reduce efficiency. It gives efficiency credibility.

It does not suppress innovation. It brings innovation into order.

It does not negate AI. It allows AI to be used at greater scale, at a higher level, and with greater safety.

Without verifiable thinking, the more powerful AI becomes, the more uneasy society becomes.

With verifiable thinking, the more powerful AI becomes, the more society can transform it into productive power.

VIII. Beyond Verifiable Thinking, What Else Can Make Us Confident in Using AI?

Let us ask: apart from verifiable thinking, what tools can truly make human beings confident in using AI? The answer is not that there are none. Law, ethics, regulation, open source, and human oversight are all important. But they are not substitutes for verifiable thinking; they must be activated by verifiable thinking.

Law is certainly important. But if law cannot restore facts, replay processes, and trace responsibility, it can only pursue accountability slowly after losses have already occurred.

Ethics is certainly important. But if ethics cannot be implemented as inspectable boundaries, records, and constraints, it easily becomes a slogan.

Regulation is certainly important. But if regulators cannot see AI’s authorization, calls, execution, and iteration process, regulation can only remain at the level of after-the-fact punishment.

Open source is certainly important. But open source does not automatically equal verification. Code can be public while the execution process remains invisible. Models can be open while specific outputs remain wrong. Rules can be written down while actual operations still deviate from the rules.

Human oversight is certainly important. But without logs, evidence, replay, and verification interfaces, human beings will still find it difficult to truly supervise AI.

Therefore, the question is not whether society needs law, ethics, regulation, open source, and human oversight. The question is whether these tools can enter verifiable structures. Only when they can be verified does law have a factual basis, ethics have executable boundaries, regulation have a handle, open source have credibility, and human oversight have real capability.

This is the fundamental position of verifiable thinking. It is not one ordinary tool among many. It is the underlying method that enables other tools to function. Without verifiable thinking, law easily becomes after-the-fact accountability, ethics easily becomes slogan, regulation easily remains punishment, open source easily remains mere code disclosure, and human oversight may fail because it cannot see the facts.

Without verification, all tools eventually retreat back into trust. In the AI age, the most dangerous thing is precisely to continue trusting a powerful system that cannot be verified.

IX. The Future Belongs to Those Who Can Build Verifiable Structures

The AI age does not lack models, computing power, applications, or conceptual hype. What is truly scarce is the thinking capable of placing AI into the real world, real responsibility, and real institutions.

Whoever can transform AI capability into verifiable execution can build a new industrial foundation.

Whoever can transform AI judgment into verifiable cognition can build a new knowledge order.

Whoever can transform AI power into verifiable governance can build a new form of institutional credit.

Whoever can connect AI with financial facts, legal responsibility, public services, and social trust truly understands the deep transformation of the AI age.

Verifiable thinking matters because it is not a technical detail. It is a method of the age.

It requires human beings, when facing any powerful system, to move beyond trust, worship, or fear, and instead ask:

Can the facts be verified?

Can the process be replayed?

Can responsibility be traced?

Can power be constrained?

Can credit be built on inspectable facts?

These questions will become the most basic questions of the AI age.

Conclusion: The Dragon Has Been Released; It Must Have a Lock

The dragon of AI has already been released.

It will not stop because human beings hesitate, nor will it wait because old institutions are uncomfortable. It will continue to enter industry, finance, law, education, governance, and the daily life of every person.

The real question now is not whether we want AI, but what kind of AI age we want.

If the AI age is built on unverifiable judgment, untraceable execution, unauditable power, and responsibility that cannot be borne, then this dragon will eventually turn back against human society.

If the AI age is built on verifiable thinking, AI’s capabilities can be released, AI’s risks can be constrained, AI’s credibility can be established, and AI’s power can enter into order.

Verifiable thinking is the dragon-taming lock of the AI age.

What it locks down is not the dragon’s flight, but the dragon’s loss of control.

What it locks down is not technological progress, but hallucination, abuse of power, and responsibility evasion.

The true AI age is not an age of trusting machines. It is an age of verifying machines, mastering machines, and making machines serve the human order.

Chapter 11 How to Build a Verifiable Structure for AI

Six Core Layers and Three Modes of Verification

In the previous articles, we moved from verifiable finance to verifiable execution, and then further to verifiable cognition and verifiable thinking. Verifiable finance asks whether key financial facts can be continuously verified. Verifiable execution asks how evidence should be preserved once AI enters the realm of action. Verifiable cognition asks how human beings can judge AI’s views, boundaries, and errors. The essay “Verifiable Thinking Is the Dragon-Binding Lock of the AI Era” discussed, at the level of thinking itself, why verifiable thinking is necessary for the development of AI. This article no longer repeats the general discussion of AI risk. It moves the question to the structural level: what should a verifiable structure for AI include, how do modes of verification differ across scenarios, and why the purpose of verification is not to prove that AI results are always correct, but to make key processes visible, reviewable, and accountable.

I. This Chapter Does Not Repeat the Risk Judgment, but Answers the Structural Question

Previous chapters have already explained, from the perspectives of responsibility, boundaries, and the dragon-taming lock, that after AI enters high-responsibility scenarios, it cannot merely be trusted; it must be verified. Repeating that AI can make mistakes no longer adds much. The question this chapter must answer is more concrete: if AI is to be verified, which structural layers should be verified?

Therefore, the focus of this chapter is not to argue once again why AI is dangerous, but to break verifiable structure into operable layers: input facts, authorization boundaries, process logging, result verification, replay audit, and responsibility return. In this way, Chapter 8 focuses on why execution facts must leave evidence; Chapter 10 focuses on verifiable thinking as an institutional constraint; and this chapter focuses on how to build a structure that is reviewable, replayable, and accountable.

II. A Verifiable Structure Does Not Require Fully Opening the AI Black Box

To build a verifiable structure for AI does not mean requiring all models to be fully open source, nor does it mean requiring ordinary users to understand every internal parameter of a model.

This is an important distinction.

Open source has value, but open source is not the same as verification. Even if a model is open source, ordinary users may not have the ability to examine it. Even if the code is public, the data, prompts, permissions, plugins, tool calls, execution records, and external environment at runtime may still remain invisible. Conversely, a non-open-source system can still form a strong verifiable structure if its key facts, authorizations, executions, and results can be recorded, reviewed, replayed, and audited.

Therefore, the focus of a verifiable AI structure is not to make every person understand the internals of the model, but to ensure that all important actions have an inspectable evidence chain.

The fact that a black box cannot be completely opened does not mean it cannot be verified. To passengers, an aircraft engine is also a black box, but the aviation system has records, inspections, standards, and accountability chains. A bank’s internal system is also a black box to customers, but account balances, transaction records, assets and liabilities, audit reports, and regulatory records must be capable of verification. AI should be treated in the same way.

We do not necessarily need to verify every internal parameter of AI. But we must verify the key facts before and after AI enters the real world:

What data did it use?
What rules did it rely on?
What authorization did it receive?
What action did it execute?
What state did it change?
What result did it produce?
Who is responsible if something goes wrong?

This is the core of a verifiable structure for AI.

III. A Verifiable Structure for AI Requires at Least Six Layers

First, the input-fact layer.

AI judgment depends on input. If the input facts themselves are not trustworthy, the stronger AI becomes, the faster errors will spread. False ledgers in finance, false evidence in law, false examination data in medicine, and false statistics in public governance can all be amplified by AI.

Therefore, AI systems first require verifiable input facts. Data source, collection time, provider, version changes, modification records, and authorization status should all be recorded. Important data should not merely be read by AI; it should be capable of review.

This is also the lesson of verifiable finance. For financial AI to make judgments, the premise is not that the model is smarter, but that ledgers, balances, reserves, liabilities, transactions, and responsibility facts are trustworthy. If the underlying facts cannot be verified, AI merely processes untrustworthy facts more quickly.

Second, the authorization-boundary layer.

AI does not automatically acquire execution authority merely because it is capable of executing. Capability is not authorization.

Important human actions require authorization. Bank transfers require authorization. Contract signing requires authorization. Medical surgery requires authorization. Corporate approvals require authorization. Government actions also require authorization. Once AI enters these scenarios, authorization boundaries become even more necessary.

AI authorization should be scope-based, conditional, version-based, revocable, and verifiable. What it may do and may not do; what amount it may execute automatically and what amount requires confirmation; under what conditions it may call external tools and what actions must be paused; and whether an original authorization remains valid after a model or strategy version changes—all these should be clearly recorded.

Complex tasks should not be fully delegated to AI without limits. Low-risk actions may be executed automatically with records preserved. Medium- and high-risk actions must be confirmed, reviewed, or reauthorized. Authorization is not a one-time click of consent. It should be a dynamic, layered, and traceable process.

Third, the process-logging layer.

AI output should not be judged only by the final result. The process also matters.

In ordinary software, execution paths are usually relatively fixed. AI’s path of judgment, however, is uncertain. It may call materials, generate intermediate judgments, use tools, compare options, adjust strategies, trigger external interfaces, and finally produce a result. If the process is not recorded, it will be difficult to know where an error occurred after the fact.

Therefore, key AI processes must leave records. These include the data version used, the prompt or task instruction, the model version, tool calls, permission checks, key intermediate conclusions, human confirmation points, returned results from external interfaces, anomaly notices, risk warnings, and final execution actions.

Not every process needs to be public, but high-risk processes must be auditable. Raw details may be inspected internally, by auditors, or by regulators. Key facts should be presented to users, customers, or the public in an appropriate form.

Fourth, the result-verification layer.

AI results should not be accepted merely because the language is fluent. So-called result verification does not mean proving that an AI conclusion is necessarily correct. It means requiring the result to be traceable back to its sources, basis, process, and accountability chain.

In finance, AI-generated asset-liability reconciliation should be traceable back to ledgers, transactions, balances, and reserve proofs. In law, AI-generated opinions should be traceable back to evidence, legal provisions, and reasoning paths. In medicine, AI recommendations should be traceable back to examination data, diagnostic grounds, and applicable boundaries. In public services, AI decisions should be traceable back to rules, application materials, and approval records.

Result verification does not require every user to become an expert. It requires the system to provide inspectable interfaces. Experts can verify in depth. Ordinary users can see key grounds. Regulators can audit the complete chain. Accountable parties can bear the final consequences.

Without this kind of verification, AI output merely appears correct. With it, even if AI output remains disputable, it can at least be inspected, reviewed, corrected, and then enter institutional credit.

Fifth, the replay-audit layer.

When AI makes a mistake, the most important task is not to argue about the result again, but to replay what happened at the time.

Who initiated the task?
What instruction did AI receive?
What data did the system read?
What was the model version at the time?
Was the permission valid?
Were there any risk warnings?
Was there human confirmation?
What actions were executed?
Which accounts, contracts, rights, or responsibilities were affected by the result?

If these questions cannot be answered, AI systems cannot truly perform social functions.

Replayability is a key part of a verifiable AI structure. Financial systems need audit replay. Legal systems need evidentiary replay. Medical systems need replay of diagnosis and treatment processes. Public governance needs replay of administrative processes. The deeper AI participates, the more important replay mechanisms become.

Future AI systems should not merely generate answers. They should generate verifiable execution records. These records do not have to be fully public, but they must be auditable, reviewable, and accountable when necessary.

Sixth, the accountability-chain layer.

AI currently cannot bear final legal responsibility. It has no life, no property, no legal personality, and cannot be punished like a natural person or a legal entity. The real responsible parties remain developers, deployers, operators, users, authorizers, and relevant institutions.

Therefore, a verifiable AI structure must ultimately lead to an accountability chain.

After an AI action occurs, it must be possible to answer: who designed the system, who provided the model, who configured the rules, who connected the data, who granted the permission, who triggered the task, who approved the execution, who benefited from the result, and who should bear the consequences.

If the accountability chain cannot be traced, AI will become a tool for evading responsibility. Institutions may say, “This was automatically generated by the system.” Individuals may say, “This was AI’s recommendation.” Platforms may say, “The user chose to use it.” In that case, AI is not improving social efficiency; it is creating an accountability vacuum.

The purpose of a verifiable structure is to prevent AI from creating such a vacuum.

IV. AI Verification Is Not a Single Technology, but an Institutional Structure

Many people easily understand AI verification as a particular technical tool: model evaluation, code audit, output detection, watermarking, logging systems, blockchain attestation, hash proofs, or third-party certification. All of these have value, but none of them is the whole structure.

A true verifiable AI structure is not a single technology. It is an institutional structure.

It includes at least structures of facts, authorization, execution, evidence, audit, and accountability. Technology is merely the tool for implementing these structures. Without institutional design, technology becomes fragmented. Without technical support, institutions become empty.

For nontechnical readers, the terms may be briefly explained as follows. A Merkle tree can compress a large number of records into a verifiable summary; if any record is changed, the summary changes. A Trusted Execution Environment (TEE) emphasizes execution within a protected hardware environment and provides external proof that it has run according to the rules. A zero-knowledge proof allows one to prove that a computation or state satisfies certain conditions without disclosing all of the underlying raw data. These technologies have different names, but in this book they occupy the same position: they are not ends in themselves, but tools that help responsibility facts, execution processes, and key states obtain a smaller, clearer, and more auditable root of trust.

This is exactly where verifiable finance can offer insight for the AI era.

The core of verifiable finance is not to move all finance onto a blockchain, nor to worship any single technology. It is to move financial credit from “trusting institutional promises” to “verifying key facts.” A transparent bank does not eliminate the bank; it makes it impossible for the bank to lie about key facts. A transparent stablecoin does not eliminate the issuer; it makes reserves, liabilities, issuance, redemption, and responsibility facts continuously verifiable.

AI is the same.

A verifiable AI structure does not eliminate AI judgment, nor does it stop AI execution. It makes AI’s key inputs, key judgments, key authorizations, key executions, and key responsibilities verifiable. Only then can AI move from tool capability into institutional credit.

V. Different Risk Scenarios Require Different Strengths of Verification

Not every AI action requires the same strength of verification.

If AI writes ordinary text, translates an email, or organizes meeting notes, the risk is low and complex verification is unnecessary. But if AI participates in fund transfers, investment transactions, legal judgment, medical advice, insurance claims, loan approvals, corporate internal control, or the allocation of public resources, it must enter a stronger verifiable structure.

Future AI behavior can be stratified by risk:

Low-risk actions may be executed automatically with basic records preserved.
Medium-risk actions require user confirmation, process logging, and result review.
High-risk actions require explicit authorization, confirmation at key nodes, complete audit replay, and identified accountable parties.
Systemic-risk actions—such as financial infrastructure, public governance, judicial assistance, medical decision-making, and automated execution by large platforms—require third-party audit, regulatory interfaces, and continuous verification mechanisms.

This does not reduce AI efficiency. It enables AI to enter more important scenarios.

Without verifiable structures, AI can only remain at the auxiliary layer.
With verifiable structures, AI can enter the execution layer.
Verification is not an obstacle to AI; it is the precondition for AI to be truly accepted by society.

VI. Three Modes of Verification: Determinate Results, Visible Processes, and Inspectable Grounds

Verification differs across fields. Many facts in verifiable finance have strong determinacy: whether an account balance exists, whether assets and liabilities match, whether reserves cover issuance, whether a transaction occurred, or whether a hash corresponds to a particular ledger. This type of verification mainly answers the question: does it match?

But once AI enters law, medicine, corporate decision-making, public services, and cognitive collaboration, many questions do not have the same determinacy as a financial ledger. Whether a legal opinion is optimal, whether a medical recommendation is entirely correct, or whether a policy judgment is sufficiently grounded often cannot be reduced to a simple right-or-wrong answer. Therefore, AI verification should not be understood as proving that a conclusion is necessarily correct. It should be understood as making the process visible: the sources of facts, the rules applied, the applicable boundaries, the omitted conditions, and the distinction between facts and inferences should all be open to questioning and review.

From this, three modes of verification can be distinguished:

First, deterministic verification applies to financial ledgers, assets and liabilities, hash proofs, and transaction records. Its focus is whether the result corresponds.
Second, process verification applies to AI execution, approvals, automated operations, and agent behavior. Its focus is whether authorization is valid, whether the process is visible, and whether the action is compliant.
Third, cognitive verification applies to AI judgment, analysis, writing, legal opinions, medical recommendations, and policy reasoning. Its focus is whether the grounds are sufficient, whether the boundaries are clear, and whether the inferences can be reviewed.

This distinction is very important. If all verification is understood as financial-style deterministic verification, people will mistakenly believe that AI cannot be verified unless it can produce absolutely correct results. In reality, AI’s uncertainty cannot be fully eliminated, but it can be constrained, observed, recorded, reviewed, and corrected. Financial facts pursue determinate results; execution behavior pursues visible processes; cognitive judgment pursues inspectable grounds and clear boundaries.

VII. Who Verifies AI?

AI verification is not completed by a single subject.

Ordinary users can verify key results and basic grounds. Professionals can verify reasoning paths and applicable boundaries. Institutions can verify authorization, procedure, and responsibility. Auditors can verify records, logs, and execution replay. Regulators can verify systemic risks. The market can verify the credit of systems through long-term use.

Different subjects verify different layers; only then can a complete structure take shape.

Ordinary people do not need to understand every model parameter, but they must have the right to know key facts. Experts do not need to control every system, but they must be able to inspect professional boundaries. Regulators do not need to interfere with every AI output, but they must be able to inspect the records and accountability chains of high-risk systems.

This is the difference between a verifiable structure and a traditional trust structure.

Traditional trust structures often require users to trust institutions, experts, platforms, and systems. The AI era cannot rely only on such trust. AI operates too quickly, affects too widely, hides errors too deeply, and makes responsibility too easy to transfer.

The new credit of the AI era should come from verification.

VIII. Without Verifiable Structures, AI Will Expand Social Uncertainty

If AI lacks a verifiable structure, society will face several serious consequences.

First, errors will become harder to discover. AI has strong expressive capability, and errors may be packaged as rigorous conclusions.

Second, responsibility will become easier to transfer. People can blame AI for errors, institutions can blame systems, and platforms can blame users.

Third, power will become more hidden. AI can influence approval, recommendation, pricing, ranking, credit granting, punishment, and resource allocation within invisible processes.

Fourth, ordinary people will become more vulnerable. When facing complex AI systems, individuals without verification rights can only accept results and cannot effectively appeal.

Fifth, institutional credit will be eroded. Society may appear more automated and efficient on the surface, while in reality becoming harder to hold accountable, harder to correct, and harder to supervise.

Therefore, building a verifiable structure for AI is not merely technical optimization. It is foundational institutional engineering for the AI era. This is a vast market: AI capability is growing rapidly, while verifiable structures remain largely blank. We cannot turn all AI uncertainty into certainty, but we can embed verification structures into AI applications and refine better structural details from real-world scenarios. The easiest to structure is verifiable execution; the area most in need of methodology is verifiable cognition; and the field that can most readily produce determinate results remains verifiable finance.

IX. Conclusion: The Stronger AI Becomes, the More It Needs Verification

The development of AI will not stop. Models will become stronger, tools will become more numerous, automation will deepen, and AI will enter the real world more extensively. Stopping AI is not the answer. Blindly trusting AI is not the answer either.

The real answer is to build verifiable structures for AI.

AI can judge, but the basis of judgment must be inspectable.
AI can execute, but execution authority must be verifiable.
AI can generate results, but results must be reviewable.
AI can participate in complex systems, but processes must be replayable.
AI can improve efficiency, but responsibility must not disappear.

From verifiable finance to verifiable execution, and then to verifiable cognition, the same logic of the era runs through all of them: human society is moving from trusting promises to verifying facts.

The most important credit of the AI era will not come from a model claiming that it is correct, nor from an institution promising that it is reliable. It will come from the fact that the structure itself can be verified.

Whoever takes the lead in building verifiable structures for AI will gain true institutional credit in the next stage.

Chapter 12 Human Beings Need the Right to Verify AI

From Passive Trust in the Age of Credit to Active Review in the Age of Verification

Core thesis: One of the most important institutional questions of the AI era is whether human beings, when facing AI judgment, AI execution, and AI power, still retain the right to verify. Whoever bears the consequences of AI results should have the right to verify AI. Without the right to verify, AI will push people back into a position of passive trust; with the right to verify, AI can become a tool that is usable, reviewable, and accountable. This article is the natural continuation of the previous series of discussions.

I. AI’s Uncertainty Requires Human Beings to Be Able to Verify AI

In terms of how it operates, AI, like human beings, has an uncertain side. Any uncertain system can make mistakes, just as human beings can make mistakes. To prevent and correct errors, human beings have built institutions, including law. Human beings clearly understand that mistakes have consequences. The question is: do we know when AI makes mistakes? After AI makes a mistake, do human beings know what facts it relied on, what process it went through, what boundaries it crossed, who authorized it, and who should bear responsibility?

If a system can affect the interests of others yet does not allow verification; if it can form judgments without disclosing its basis; if it can execute automatically without leaving a process trail; if it can produce consequences without leaving an accountable subject, then it is no longer merely a technical tool. It becomes a new kind of unaccountable power. In extreme cases, such power may approach a totalitarian structure; the degree of totalitarianization depends on the scope within which it escapes accountability. Without a verification structure, AI’s “right to explanation” may also degenerate into a more fluent form of propaganda. The AI era must therefore raise a right that is more fundamental and more complete than the “right to explanation”: human beings need the right to verify AI.

II. In the Age of Credit, Human Beings Were Passive Believers

Traditional society was built on credit. This kind of credit is not without value. Modern society cannot require every individual to personally verify every fact. When a bank says an account balance is what it is, users usually can only believe it. When an insurance company refuses to pay a claim, the customer often finds it difficult to know the full basis. When a platform says an account has violated rules, the user often sees only a brief notice. When a recruitment system filters out a candidate, the candidate may not know whether the result was caused by an error in resume facts, algorithmic bias, or some invisible standard. When a public-service system denies eligibility, ordinary people may also be unable to review the underlying chain of facts at low cost.

This is the basic structure of the age of credit: institutions control facts, rules, and judgment processes, while individuals accept results. Individuals may sometimes appeal, sue, or complain, but these mechanisms are costly and slow, and they often occur only after the result has already been produced.

Without institutional credit, complex society cannot operate. But the consequence is this: because many key facts cannot be verified, they can only be believed; to a large extent, the foundation of the social edifice is belief.

Verifiable finance proposes a move “from trusting institutional promises to verifying financial facts.” In essence, this is a breakthrough against the structure of passive trust. It does not seek to eliminate institutions, but to ensure that key facts cannot be stated only by institutions unilaterally. Now, the same logic must be extended to AI.

III. AI Superimposes Institutional Black Boxes, Algorithmic Black Boxes, and Execution Black Boxes

After the emergence of AI, the problem becomes more complex.

In the past, institutional black boxes at least involved people who could be questioned, documents that could be obtained, and responsible persons who could be pursued. Although the cost was high, a path of accountability still existed in theory. Traditional software systems also had algorithms and rules, but many processes were relatively fixed, and the relationship among input, rule, and output was easier to audit.

AI is different. AI judgment is probabilistic, its output is uncertain, and its process is jointly affected by the model, context, prompts, data sources, tool calls, and system settings. When AI is embedded in banking, insurance, healthcare, recruitment, platform governance, judicial assistance, and public-service systems, individuals no longer face only institutions, nor only algorithms. They face the superimposition of institutional power, algorithmic judgment, and automated execution.

This produces a new black-box structure: the institution says the system made the judgment, the system says the model produced the output, and the model process is difficult to explain. In the end, the affected person sees only a result and cannot know how that result was produced.

Without the right to verify, AI will make passive trust even more severe. In the past, people trusted institutions; now people may be forced to trust “institution + algorithm + model + automated execution system.” This is not progress in credit. It is a regression of credit.

True AI progress should not make it harder for people to question. It should make it easier for them to review. The stronger AI becomes, the more verification structures it needs; the more automated AI execution becomes, the less responsibility can be allowed to disappear into technical systems.

IV. The Right to Verify Is Not Merely the Right to an Explanation; It Is the Right to Active Review

Today, discussions of AI governance often mention the “right to an explanation.” This right is, of course, important. Users have the right to know why AI produced a certain result, and regulators also need to understand the basic reasons behind automated decisions.

But the right to an explanation is not the same as the right to verify.

The right to an explanation tends toward “the system stating its reasons.” It answers questions such as: why was this judgment made? What rules did it rely on? Which major factors were used?

The right to verify goes further. It asks: Is this fact true? Is this data correct? Did this model operate within its authorized scope? Was this process recorded? Can this result be replayed? Did this decision exceed its authority? If damage was caused, who is responsible? Users have the right to demand that a system provide a basis that can be reviewed, especially in paid services and high-risk services.

An explanation may be given unilaterally by the system; verification must allow users, third parties, regulators, or responsible institutions to conduct review. An explanation may sound reasonable; verification requires that people be able to check whether it is actually reasonable.

Therefore, what the AI era needs is not to stop at a mere right to explanation, but to move toward a more fundamental right to verify. The right to explanation is a precondition and component of the right to verify, but it cannot replace the right to verify. Without a verification structure, explanation may merely become more fluent persuasion; with a verification structure, explanation can become a trustworthy account.

V. Whoever Bears the Consequences Should Have the Right to Verify

The legitimacy of the right to verify comes from a simple principle: whoever bears the consequences should have the right to review those consequences.

If AI judgment affects a person’s loan, insurance, medical care, admission, employment, transaction, assets, credit, platform account, eligibility for public services, or legal position, then that person should not be merely the bearer of the result. He or she should also be a verifier of the key facts and key processes.

Where there are no consequences, verification standards can be lowered, or verification may not be needed. The more serious the consequences, the stronger the right to verify. The more irreversible the impact, the stricter the verification obligation.

This is not anti-AI, nor does it require every ordinary user to understand the internal structure of a model. On the contrary, it is the institutional precondition for allowing AI to enter high-risk fields. Without the right to verify, the stronger AI becomes, the more anxious society becomes. With the right to verify, the stronger AI becomes, the more safely it can be used.

VI. The Right to Verify Does Not Require Opening AI’s Entire Internal Black Box

Once the right to verify is proposed, an objection immediately arises: AI is internally complex; model weights, training data, reasoning paths, trade secrets, and safety mechanisms cannot all be disclosed, and ordinary people cannot understand them all.

This objection is partly correct, but it misses the core issue.

The human right to verify does not require opening AI’s entire internal black box, nor does it require every user to understand how a neural network forms a particular intermediate representation. What the right to verify requires is this: whenever AI enters real-world execution, affects important human interests, and produces legal or economic consequences, the key external facts must be verifiable, recordable, replayable, and accountable.

The inside of a model may be complex, but the boundary of authorization must be clear. Algorithms may involve trade secrets, but the key data used, the applicable rules, whether human review occurred, and whether an appeal channel exists cannot be entirely invisible. AI reasoning may not be unique, but the facts it relied on, the tools it called, the records it formed, and the results it produced must be capable of being inspected when necessary.

Verifiable finance does not make everything public to everyone. It ensures that key financial facts cannot be hidden, tampered with, or rewritten after the fact. The AI right to verify is the same. It is not an unlimited right of disclosure, but a right to review key facts; it is not a right to pry into trade secrets, but a safeguard against unaccountable automated power.

VII. Human AI Verification Rights Should Include at Least Seven Elements

First, the right to know. People should know whether AI was used in an important judgment or execution, and what role AI played: whether it merely assisted, offered advice, performed automatic scoring, or executed directly.

Second, the right to verify authorization. Whenever AI executes important matters on behalf of a person, especially matters involving finance, contracts, accounts, identity, healthcare, and law, authorization must be clear, limited, and revocable. Users should be able to verify who authorized it, to whom authorization was given, what was authorized, how long the authorization lasts, where the boundary of authority lies, and whether automatic expansion of authority occurred.

Third, the right to review facts. The key facts on which AI judgment relies should be reviewable by the affected person. Whether personal data is wrong, transaction records are complete, medical materials are missing, credit records have expired, or the facts underlying platform punishment are true — none of these can be covered over by the sentence “the model has already judged.”

Fourth, the right to replay the process. High-risk AI execution cannot provide only a result. When necessary, the key process should be replayable: what the task was, what the input was, what materials were called, what tools were used, what confirmations were made, whether exceptions were triggered, and what the final output was.

Fifth, the right to boundary explanation. AI should explain its scope of application and uncertainty. Is it only a probabilistic judgment? Does it require expert review? Does it rely on incomplete materials? Has it exceeded its training or application scenario? Such boundary information is itself part of verification.

Sixth, the right to human review. When major rights and interests are affected, people should not be forced to face only a machine conclusion. There should be a human or legal-person review mechanism that can understand responsibility, make a new judgment, and bear the consequences.

Seventh, the right to seek responsibility. The ultimate purpose of verification is not curiosity, but accountability. AI is not the final responsible subject. The institution deploying AI, the subject authorizing AI, the organization designing the rules, and the person or legal entity using AI to make decisions must all return to the chain of responsibility.

Together, these seven elements constitute the basic framework of human AI verification rights.

VIII. Judgment and Verification: Final Judgment Must Not Be Completely Handed Over to Machines

Here we must also distinguish judgment from verification.

Judgment is not entirely the same as verification. Judgment is an integrated activity that includes factual judgment, value judgment, risk judgment, responsibility judgment, and trade-off judgment. Verification places more emphasis on reviewing facts, processes, bases, boundaries, and responsibility.

But the right to verify is the precondition for judgment. Without the right to verify, human beings do not have sufficient materials with which to make judgments. Without facts, processes, and boundaries, so-called “consent” may be nothing more than passive acceptance.

AI can participate in judgment, and in many local judgments it may even surpass ordinary people. AI can help doctors read medical images, help lawyers search cases, help financial institutions identify risks, help enterprises optimize processes, and help individuals organize complex information. But AI cannot therefore obtain a monopoly over final judgment.

The higher the risk of a scenario, the less final judgment can be completely handed over to machines. Final judgment involves not only computational efficiency, but also consequence-bearing, rights protection, value ordering, and responsibility allocation. AI can provide recommendations, but human beings or legal persons must bear responsibility. AI can assist verification, but it cannot eliminate the human position of review. AI can improve judgment efficiency, but it cannot deprive people of the right to question the basis.

Therefore, the mature structure of the AI era is not “machines judge, humans obey,” but “machines participate in judgment, humans retain verification; machines improve efficiency, responsibility returns to human beings or legal persons.”

IX. Different Scenarios Require Different Levels of Verification Rights

The right to verify cannot be designed as a one-size-fits-all rule. Low-risk, medium-risk, and high-risk scenarios should have different standards.

Ordinary chatting, writing suggestions, daily translation, and general information organization usually do not require the mandatory generation of a complete evidence package. Users can rely on common sense, experience, and subsequent checking to make their own judgments.

Medium-risk scenarios, such as professional consulting, business analysis, contract drafts, learning assessments, and internal enterprise management suggestions, should at least preserve the task, sources of materials, versions, key assumptions, and modification records. The focus here is to allow users to know AI’s basis and boundaries.

High-risk scenarios, such as financial transactions, insurance claims, medical advice, legal judgments, recruitment screening, credit scoring, platform punishment, allocation of public services, asset disposal, and administrative auxiliary decision-making, must have stricter verification obligations. Systems should generate accessible records and preserve authorization paths, key facts, model or rule identifiers, human review status, appeal channels, and responsible subjects.

The highest-risk scenarios, especially those involving fund transfers, personal safety, legal consequences, public power, and irreversible losses, should require mandatory verification, mandatory record-keeping, mandatory human review, or mandatory third-party audit.

Such stratification ensures that verification does not become an infinite cost, and also prevents AI from entering an unaccountable state in the name of efficiency. The right to verify is not about increasing social transaction costs without limit. It is about using automated records, automated auditing, and necessary human review to reduce long-term trust costs.

X. The Right to Verify Must Be Productized, Standardized, and Legalized

If the right to verify remains only an idea, it will be difficult for it to function. A truly useful right to verify must be productized, standardized, and legalized at the same time.

Productization means providing clear verification entry points in AI systems. Ordinary users do not need to read long logs every day, but next to key results there should be functions such as “view basis,” “view authorization,” “view process,” “request human review,” and “export verification record.” The right to verify should exist like a button, not hide inside long terms and conditions.

Standardization means enabling AI systems to generate verification records in a unified format. Such records should include a task summary, model or system version, key inputs, main sources of materials, authorization records, tool calls, important intermediate results, final output, user confirmations, exception alerts, timestamps, signatures, or hashes. Different scenarios may use different templates, but the basic structure should be readable, exportable, auditable, and tamper-resistant.

Legalization means turning the right to verify from a voluntary corporate service into a basic obligation in high-risk scenarios. The law need not require the disclosure of all model details, but it should require automated decisions to preserve key records, major consequences to provide review mechanisms, regulators to be able to audit key facts under lawful procedures, and affected persons to have the right to obtain necessary verification materials.

Technology provides capability; products provide entry points; standards provide formats; law provides enforceability. Only when all four are combined can the right to verify truly take root. Technically, different combinations may be used, including log signatures, timestamps, hash-based evidence preservation, third-party audits, blockchain anchoring, verifiable computation, and zero-knowledge proofs. The point is not to pile up technical terms, but to ensure that key processes can be independently verified when necessary.

XI. The Right to Verify Is Also an Institutional Constraint on the Technological Aristocracy

The greatest danger in the AI era is not only AI itself losing control, but also the people and institutions that control AI escaping social constraint.

A small number of institutions possess models, computing power, data, capital, channels, and automated execution capabilities. If their judgments cannot be verified, their execution cannot be replayed, their gains cannot be explained, and responsibility cannot be pursued, then they may form a new structure of technological power. Ordinary people will not be facing a tool, but a machine-like power that cannot be seen, questioned, clarified, or held accountable.

This is where the significance of the right to verify lies. It is not anti-technology. It opposes unaccountable technological power. It is not anti-AI. It opposes the deprivation of the human position of review in the name of AI. It does not protect backwardness. It builds institutional boundaries for advanced productive forces.

Verifiable finance constrains financial power. Verifiable execution constrains AI action. Verifiable cognition constrains machine judgment. The human right to verify brings all of these constraints back to specific human beings.

Without the human right to verify, verifiable structures may become merely internal corporate audit tools. With the human right to verify, verifiable structures become institutional capabilities for protecting users, consumers, citizens, patients, investors, and workers.

XII. The Difference Between Human Verification and Machine Verification

Ensuring safety and reliability through technology is the prevailing approach of today’s AI companies. It tries to convert uncertain problems as much as possible into deterministic problems. This is effective locally, especially for internal security, permission control, log auditing, and format checks. Machines can verify whether a signature is correct, whether a hash is consistent, whether a log is continuous, whether permissions match, whether a timestamp exists, and whether a data format is compliant. They are suited to formalized, rule-based, computable problems. They are fast, low-cost, and highly repeatable. Determinism is very important, but it is not everything. The greatest paradox here is that uncertainty is precisely an important feature of generative AI.

Human verification has externality. Human beings do not merely verify whether a computational result holds. They also ask whether that result should be applied to a particular person, a particular scenario, and a particular consequence. Machines can check whether rules have been executed. Human beings must further examine whether the rules are legitimate, whether authorization is real, whether the consequences are bearable, and whether someone is responsible.

The former is more technical; the latter is more institutional. The former can replace part of manual checking; the latter cannot be completely replaced by machines.

In the AI era, we certainly need machines to verify machines. The output of one AI can be checked by another AI for facts, contradictions, and risks. Systems can also automatically generate logs, evidence packages, and audit reports. This can reduce the cost of human verification and improve audit efficiency. But machine verification cannot become the final closed loop. In the end, human beings or legal persons must still decide whether to accept the conclusion, whether to allow continued execution, and whether to bear the corresponding responsibility.

XIII. With Respect to AI, What Else Must Human Beings Retain Besides the Right to Verify?

With respect to AI, human beings must retain not only the right to verify, but also the right to authorize, the right to refuse, the right to correct, and the right of final judgment. The right to verify is not an isolated right; it is the foundation that makes these other rights possible. Without verification, there is no real authorization. Without verification, there is no effective refusal. Without verification, there is no correction or accountability.

This is also an extension of the problem that machines cannot bear responsibility. AI can execute tasks, offer judgments, and generate proposals, but it has no life, no property, and no independent capacity for responsibility. It cannot bear legal consequences in the way a human being or legal person can. A machine cannot truly be punished, nor can it bear moral remorse.

Therefore, the basic relationship between human beings and AI cannot be designed as “machines decide, humans accept.” It should be designed as “humans authorize, AI executes; humans verify, responsibility returns to human beings or legal persons.” Authorization determines whether AI may act; verification determines whether what AI has done is trustworthy; responsibility determines who bears the consequences.

From this perspective, the right to verify is not an accessory in the human-AI relationship. It is the central axis of that relationship. Without the right to verify, authorization becomes surrender of power, automation becomes loss of control, and responsibility is absorbed by the technical system. With the right to verify, AI can retain its nature as a tool, and the human subject position will not be swallowed by automated processes.

In the AI era, the right to verify is a basic human right, like the right to life and the right to survival. However prominent one may appear in the real world, without the right to verify, one is still a slave to AI.

XIV. From Verifiable Finance to the Human Right to Verify: A Natural Theoretical Progression

Verifiable finance first addresses the problem of financial credit: key financial facts cannot rely solely on institutional promises; they must enter a verifiable structure. This is deterministic verification, and AI also needs it.

Verifiable execution further addresses the problem of AI action: AI cannot provide only results; key execution processes must be recordable, replayable, auditable, and accountable. This is process verification.

Verifiable cognition addresses the problem of AI judgment: we cannot simply believe the fluent conclusions produced by machines. We must examine factual bases, reasoning frameworks, uncertainty, and boundaries of application. This is judgment verification.

The previous articles have discussed these three questions. The human right to verify pushes these questions to the level of rights. It asks: who has the right to initiate verification? Who has the right to obtain records? Who has the right to demand review? Who has the right to refuse unverifiable AI decisions? Who has the right to bring machine conclusions back into a human responsibility structure? This is a basic right of human beings in the AI era. Only when responsibility returns does the whole logic truly close its loop.

The true AI era will not be defined merely by institutions having stronger audit systems, nor by AI companies having more polished safety reports. It will be defined by ordinary people, when affected by AI, being able to ask the most basic questions: What is your basis? How did you reach this result? How can I review it? How can I appeal? Who bears responsibility?

XV. Conclusion: The Stronger AI Becomes, the More Human Beings Need the Right to Verify

AI will become stronger and stronger. This is the general trend. Once AI enters the real world, it must be subject to verification. The closer AI comes to the role of an agent, the less authorization can be vague. The more AI participates in judgment, the less human beings can give up review. The more AI can produce consequences, the more responsibility must return to human beings or legal persons.

In the age of credit, people were often required to believe. In the age of verification, people should have the right to review.

This is the most basic institutional requirement of the AI era. In the past, credit was built on institutional promises. In the future, more and more credit will be built on verifiable facts. In the past, people were recipients of results. In the future, people should become initiators of verification. In the past, machines were merely tools. In the future, even if AI appears more like a collaborator, it must still remain under the human right to verify.

AI can provide answers, but human beings must retain the right to question those answers. AI can execute tasks, but human beings must retain the right to verify that execution. AI can participate in judgment, but human beings must retain the right of final review.

In one sentence: in the AI era, human beings must not only be judged by AI; they must have the right to verify AI. Only in this way can human beings retain agency in the AI era.

Part IV From AI Being Verified to Verification-Oriented AI

Chapter 13 Three Types of Facts in the Age of AI: Determinate Facts, Procedural Facts, and Cognitive Facts

Why Different Facts Must Be Verified in Different Ways

I. Classification Is the Foundation of Solving the Verification Problem

When something new appears, the first thing that appears is often confusion.

AI is no exception. After large language models emerged, people quickly began discussing whether AI would replace human beings, whether superintelligence would arise, whether certain professions would disappear, and some even kept making sensational predictions about immortality, mind uploading, and the future world. In an attention-driven era, the more startling a claim is, the easier it is to spread; the quieter work of conceptual clarification is much less likely to be rewarded by the market.

In the early stage of AI development, if there is no overall classification, facts, opinions, predictions, execution processes, responsibility relationships, and cognitive judgments will be mixed together. Real theoretical work often does not begin with creating sensation, but with organizing confusion.

Several previous articles have already discussed the relationship between AI and verification from different angles. From Trust to Verification discusses a transformation in viewpoint; Verifiable Thinking Is the Dragon-Taming Lock of the AI Age discusses why AI must be constrained by verifiable thinking; How to Build a Verifiable Structure for AI discusses how authorization, input, process, output, replay, and responsibility should enter a verifiable structure; Human Beings Need the Right to Verify AI emphasizes that final review cannot be completely handed over to machines; Verification Is Not Prediction points out that future judgments cannot be treated as facts that have already occurred.

The further question is: what exactly should be verified in the age of AI? The answer seems simple: verify facts. But the problem is that facts are not all the same kind of fact.

If we do not first distinguish types of facts, verification becomes a vague slogan. People will place financial facts, content provenance, execution processes, model outputs, expert opinions, AI judgments, and future predictions into the same concept of “fact-checking.” Once that happens, verification methods inevitably become confused.

Therefore, we need the right classification. Only after objects are classified can methods correspond to them; different objects must enter different verification methods.

Determinate facts require factual verification.

Procedural facts require replay.

Cognitive facts require review.

These three sentences are the core of this article. Without classification, verification can only remain an attitude; with classification, verification can become a method.

II. Existing Theorists Have Seen Partial Problems

AI theorists are not blind to the problem.

Bender and her co-authors used the phrase “stochastic parrots” to remind people that large language models generate language, but this does not mean they understand the facts and meanings behind language. Marcus and Lenat pointed out that today’s generative AI can produce outputs that appear plausible, but this does not mean it can provide reliable facts. Floridi described the special position of generative AI as “agency without intelligence,” reminding us that we cannot discuss only whether AI understands; we must also discuss how AI acts. AI safety researchers such as Russell and Bengio, from different perspectives, emphasize the importance of uncertainty, supervision, evaluation, and control.

These judgments are all important. They touch, respectively, on language, facts, action, safety, and control in the age of AI. But they still enter the problem from different directions. Some see that language models do not equal understanding; some see that model output does not equal reliable knowledge; some see that AI’s capacity for action can be separated from intelligent understanding; some see that AI safety requires supervision and control.

Beyond theory, fact-checking, content provenance, AI auditing, process supervision, and multi-model evaluation have also become real directions. Automated fact-checking and claim verification research is using large language models for claim extraction, evidence retrieval, truth-value judgment, and explanation generation. Google’s ClaimReview and Fact Check Tools also show that structured checking around factual claims has entered a tooling stage.

Content provenance is also developing rapidly. C2PA and Content Credentials seek to use open standards to record the provenance and editing history of digital content, so that a piece of content’s source, generation method, and modification process can be traced. AI auditing and regulation have also begun requiring systems to have logging capabilities. For example, Article 12 of the EU AI Act requires high-risk AI systems to be technically capable of automatically recording event logs, in order to ensure traceability throughout the system lifecycle. OpenAI’s process supervision research also shows that in complex reasoning, looking only at the final result is not enough; the intermediate process steps themselves also need to be checked. LLM-as-a-Judge research further shows that AI can participate in evaluation and review, but its reliability, bias, and consistency still need to be evaluated.

These developments show that the verification problem in the age of AI is no longer an abstract idea. It is entering tools, standards, regulation, and research. But they also expose a more basic confusion: people still discuss objects of different natures under the same concept of “fact-checking.” Factual claims, content provenance, execution processes, model reasoning, expert opinions, and future predictions are obviously not the same kind of object.

Therefore, this article does not deny existing work in fact-checking, content provenance, or AI auditing. Rather, it proposes a more basic classification: facts in the age of AI should at least be divided into determinate facts, procedural facts, and cognitive facts. Different facts cannot be handled with the same verification method.

III. Determinate Facts: Structural Certainty Makes Truth and Falsity Judgable

The first category is determinate facts.

Determinate facts are facts that have already occurred, have a clear state, and can be verified through external evidence, ledger structures, signatures, records, legal documents, or third-party confirmation. Financial facts, transaction facts, reserve facts, account-balance facts, identity facts, authorization facts, and contract-signing facts all belong to this category.

The key questions for this type of fact are: Is it true or false? Does it exist? Is its state consistent? Has it been tampered with?

For example, when a stablecoin project claims that it has sufficient reserves, this is not an opinion, nor is it a prediction. It is a question of determinate fact. Whether the reserves exist, what the amount is, what the issuance volume is, whether reserves and issuance correspond, whether the bank account is real, whether audit records exist, and whether on-chain issuance and off-chain assets are consistent—all of these must be verifiable.

The way to handle determinate facts is not to ask AI to judge, “I think it is true.” It is to establish a verification structure.

Here we may borrow the principle of transparent banking: structural certainty is what makes truth and falsity judgable. A transparent bank is not trustworthy because the bank calls itself transparent. It is trustworthy because accounts, transactions, reserves, authorization, hashes, reference chains, and the public credit root form a verifiable structure, so that external parties can judge whether the accounts are consistent, whether reserves exist, and whether records have been tampered with.

Therefore, the verification method for determinate facts is factual verification.

Factual verification depends on original records, confirmation by both parties, third-party confirmation, signatures, hashes, timestamps, reference chains, account-to-account reconciliation, on-chain and off-chain consistency checks, and, when necessary, anchoring to a public credit root. AI can participate in this verification, but AI is not the source of the facts. AI can help extract claims, compare records, detect contradictions, check hashes, and generate verification reports, but the final basis must still come from structured evidence.

This distinction is very important. Fact-checking AI already exists, but it mainly handles factual claims in text. It can check whether someone said something, whether an event occurred, or whether public data supports a claim. But financial facts, reserve facts, transaction facts, and authorization facts cannot rely only on text-based fact-checking. They need a transparent-banking-style verification structure.

In other words, determinate facts are not made true by AI judgment. Their truth and falsity become judgable because of the verification structure.

IV. Procedural Facts: Without Traceability, There Can Be No Replay

The second category is procedural facts.

Procedural facts do not mainly answer the question “Is the conclusion true or false?” They answer: Did this process occur? Did it occur according to the rules? Was there overreach? Can it be traced? Can it be replayed?

AI execution processes, approval processes, tool calls, data calls, human confirmation nodes, risk-control triggers, platform penalties, insurance claims, recruitment screening, medical assistance, and legal-document generation may all contain procedural facts.

For example, when an AI financial agent performs a task on behalf of a user, the issue is not only what the final result was. It also includes: Was the user authorized? What was the scope of authorization? What data did the AI call? Which model version was used? Did it call external tools? Did it exceed its authority? Did it trigger risk controls? Was there human confirmation? When an exception occurred, was execution paused? Who approved the final execution?

These are not simple true-or-false judgments. They are questions of process replay.

Therefore, the verification method for procedural facts is replay.

Replay depends on traceability. AI systems themselves must allow users, institutions, and regulators to leave traces. Without traces, there can be no replay; without replay, there can be no process verification.

This kind of trace is not an ordinary internal log. If ordinary logs are kept entirely by the platform itself, there remains the problem of later modification, selective presentation, and monopoly over interpretation. Stronger traceability should include key inputs, authorization scope, model version, tool calls, human confirmation, output results, exception nodes, and execution status, and should hash key nodes and key content. When necessary, these hashes can be anchored to a public credit root, so that it can later be proven that the process records were not tampered with. Of course, the strength of traceability should be jointly determined by user requirements, risk level, and privacy-related restrictions on calls.

In general, procedural facts also require layered treatment. Core links that are high-risk, high-value, and involve assets, rights, responsibility, or the public interest should have hard traceability and strong replay. Low-risk, high-frequency, low-value ordinary interactions may use sampling records, exception-triggered traceability, and post-event audits. A verification structure should not infinitely expand cost; it should fix the key facts and key processes that most need verification.

Therefore, procedural facts are not solved by “AI explaining itself.” They should be solved, according to user requirements and risk level, through “traceability—hashing—replay—responsibility.”

V. Cognitive Facts: Opinions Cannot Be Factually Verified; They Can Only Be Reviewed

The third category is cognitive facts.

Cognitive facts are the easiest category to misunderstand. By cognitive facts, I do not mean that opinions, evaluations, and predictions themselves are determinate facts. Rather, I mean the factual object formed when AI or humans produce cognitive conclusions: who made the judgment, what facts it relied on, what reasoning process was used, what standards were adopted, where the boundaries lie, whether opposing evidence exists, and whether the result can be evaluated afterward.

AI analysis, evaluation, risk judgment, credit scoring, investment advice, legal opinion, medical advice, recruitment judgment, policy analysis, article evaluation, and future prediction all belong to this category.

Such content cannot be directly verified as true or false in the way a bank balance can. For example, “an account had 100 million dollars at a certain time” is a determinate fact. “This project is high-risk” is a cognitive judgment. “This asset will rise in the future” is a prediction. Mixing these three together is one of the most dangerous conceptual confusions in the age of AI.

The way to handle cognitive facts is not factual verification, but review.

Review examines whether the basis is real, whether concepts are clear, whether reasoning jumps steps, whether the logic is consistent, whether key counterexamples are omitted, whether opinions are being disguised as facts, whether predictions are being disguised as facts, whether boundaries are explained, and whether different views are allowed into the discussion.

This kind of review resembles expert evaluation and seminars. One AI may generate an opinion; another AI may challenge it from an opposing perspective; a third AI may check its factual basis; a fourth AI may examine its logical structure. The value of multi-AI review is not to give another AI final judicial authority. It is to broaden the field of view, expose gaps, and improve logical consistency.

This is also the method we are using now: a human proposes the core judgment; AI helps organize structure and expression; other AIs are then used to challenge, question, and identify weaknesses from external perspectives; finally, a human still decides which criticisms stand and which do not.

Therefore, the verification formula for cognitive facts is: multi-party review plus final human judgment.

Here, “human” is not an abstract human. In fields with a high professional threshold, final review should, as far as possible, be undertaken by people with responsibility-bearing capacity and professional competence, and may introduce expert evaluation, adversarial discussion, and appeal mechanisms. Human final judgment is not arbitrary decision-making. It is the final confirmation of basis, logic, boundary, and responsibility after multi-party review.

AI can become a tool for cognitive review, but it cannot become the final source of facts. Otherwise, we are merely using one black box to examine another black box.

VI. Opinions Are Not Facts, and Predictions Are Not Facts

In the age of AI, the most easily confused problem is that opinions and predictions are packaged as facts.

Opinions are not facts.

The fact that someone expressed an opinion can itself become a determinate fact. For example, whether an expert said a particular sentence on a particular date can be checked through the original text, source, recording, and context. But the content of that opinion does not thereby become a determinate fact. If an expert says “a certain asset has no value,” the act of making that statement can be verified, but “that asset has no value” remains a cognitive judgment, not a determinate fact.

Therefore, opinions must be handled on two levels: the existence of an opinion can be verified; the content of the opinion can only be reviewed.

Predictions are also not facts.

The fact that an AI made a certain prediction at a certain time can be recorded. But the content of the prediction itself is not a fact. A prediction points to the future, and the future has not yet occurred. Only after the future event actually occurs can the result become a determinate fact that can be verified.

Therefore, predictions must also be handled in layers: the act of prediction can be recorded; the basis of prediction can be reviewed; the outcome of prediction can only be tested after future facts emerge.

This distinction is extremely important. AI is especially good at generating judgments, explanations, and predictions, but what society needs most is verifiable facts. If AI judgments are treated as facts, and AI predictions are treated as facts, machine outputs will wear the outer garment of fact, and will then influence finance, law, medicine, recruitment, credit scoring, and public decision-making.

Therefore, the factual layer in the age of AI cannot treat all content as fact. It must strictly distinguish which objects are determinate facts, which are procedural facts, and which are cognitive facts; which are merely opinions, and which are merely predictions.

Opinions and predictions can be recorded, tracked, and reviewed, but they cannot directly obtain the status of determinate facts.

VII. The Three Types of Facts Are Not Industry Categories, but Verification-Nature Categories

Here we must avoid another misunderstanding: determinate facts, procedural facts, and cognitive facts are not divided by industry. They are divided by the nature of verification.

Financial facts are often determinate facts, but financial scenarios also include procedural facts and cognitive facts. For example, when AI performs loan approval, account balances, income records, and contract documents are determinate facts. Which data the AI called, what approval steps it went through, and whether human confirmation was triggered are procedural facts. The AI’s judgment that “this person is high-risk” is a cognitive fact.

Similarly, medical scenarios may also contain all three types of facts. A patient’s laboratory test data, imaging records, and medication records are determinate facts. The AI-assisted diagnostic process, physician confirmation process, and system-prompt process are procedural facts. The diagnostic opinion, risk judgment, and treatment recommendation formed by AI or physicians are cognitive facts.

The same is true in law, recruitment, insurance, platform governance, and public services.

In the same event, the three types of facts often exist at the same time. They are not isolated from one another. Rather, their dominant verification methods differ. Determinate facts are mainly verified through factual verification; procedural facts through replay; cognitive facts through review.

The three types of facts also refer to and transform into one another. An AI risk judgment itself belongs to the category of cognitive facts. But when that judgment is used for loan rejection, platform punishment, or asset disposal, the subsequent execution process forms procedural facts. If it is later discovered that the income records, transaction data, or identity documents on which the judgment relied were wrong, the problem returns to the verification of determinate facts. The purpose of classification is not to cut the world into three unrelated pieces. It is to allow the chain of responsibility to be traced from cognitive judgment to execution process, and then back to foundational facts.

This is precisely the significance of classification.

If we do not classify, all verification will be called fact-checking. If we classify clearly, we know that different objects must enter different structures.

VIII. AI Can Participate in Verification, but It Cannot Replace the Verification Structure

AI can certainly be used for verification. It can even be said that more and more verification-oriented AI will appear in the future.

But verification-oriented AI is not the final judge. It is an auxiliary tool.

For determinate facts, AI is a verification assistant. It can extract factual claims, search for evidence, compare records, check anomalies, and generate reports. But the final basis must come from original records, signatures, ledgers, legal documents, hashes, reference chains, and the public credit root.

For procedural facts, AI is a replay assistant. It can help organize authorization chains, input chains, tool-call chains, human confirmation nodes, and exception records, so that humans and regulators can see how things happened.

For cognitive facts, AI is a review assistant. Like an opposing view in expert evaluation, it can check logic, discover gaps, add counterexamples, and point out conceptual confusion. But whether a judgment is ultimately accepted still requires humans to retain final review authority.

Therefore, AI can participate in verification, but it cannot replace the verification structure. AI can improve checking efficiency, but it cannot become the source of facts. AI can help replay processes, but it cannot replace traceability structures. AI can review opinions, but it cannot turn opinions into facts.

More importantly, verification-oriented AI itself must also be verified. Its authorization source, input data, model version, checking process, output conclusion, and exception nodes should also enter traceability, replay, and review structures. Otherwise, so-called AI verification becomes a black box stacked on top of another black box.

Conclusion: Facts in the Age of AI Must Be Reclassified

In the age of AI, facts are easily submerged in content, opinions, predictions, and processes.

When AI can generate text, images, videos, code, judgments, and suggestions at scale, society no longer faces only a shortage of information. It faces information excess, unclear facts, unclear provenance, unclear responsibility, and unclear judgment boundaries.

In such an age, verification is more important than belief. But more precisely, classification is more important than speaking vaguely about verification.

Different facts have different natures, and different natures require different methods. Determinate facts require structural verification. Procedural facts require traceability and replay. Cognitive facts require multi-party review. Opinions must not pretend to be facts. Predictions must not pretend to be facts. AI outputs cannot naturally obtain the status of facts.

A truly verifiable AI is not an AI that is always correct, nor is it another AI that judges all AI. It is a system in which the facts, processes, and judgments formed with AI participation enter the correct verification structures.

Without classification, verification can only remain an attitude; with classification, verification can become a method.

What the age of AI needs most is not more sensational prediction, but a clearer classification of facts. Only after facts are distinguished can human beings know what should be verified, how it should be verified, who should verify it, and who should be responsible for the results after verification.

Chapter 14 How an AI Responsibility Chain Is Formed

From Fact Verification and Process Replay to Cognitive Review

Introduction

After AI makes a mistake, the sentence people most often say is: “The AI was wrong.”

But this sentence actually explains nothing.

Why do people accept the phrase “the AI was wrong”? Because humans unconsciously treat AI as a basic unit similar to a “person,” and therefore stop asking further questions.

With AI, we must continue to ask: if the AI was wrong, was the underlying fact wrong, or was the execution process wrong? Was the model’s judgment wrong, or was the human authorization wrong? Was the data source problematic, or was the rule design flawed? Did the AI fail to execute the task according to instructions, or did humans hand over a judgment to AI that should not have been handed over?

This is, in essence, a question of event accountability. To understand an AI responsibility chain, we must first look at how human society holds events accountable.

When a doctor misdiagnoses a patient, we cannot simply say, “the hospital was wrong.” We examine medical records, test results, the diagnostic process, the treatment procedure, the physician’s qualifications, the hospital’s system, and what the patient was told.

When a traffic accident occurs, we cannot simply say, “the car crashed.” We examine the facts at the scene, driving behavior, traffic rules, the vehicle’s condition, traffic lights, braking records, surveillance footage, and causality.

When financial fraud occurs, we cannot simply say, “the accounts were wrong.” We examine original vouchers, the bookkeeping process, approval procedures, audit records, management responsibility, and whether anyone deliberately concealed the truth.

When a financial risk incident occurs, we cannot simply say, “the system failed.” We examine whether authorization was valid, whether the transaction was real, whether risk controls were triggered, whether approvals were compliant, whether risks were disclosed, and whether responsibility crossed its proper boundary.

In summary, the human method of accountability is not a one-sentence assignment of blame. It traces backward along facts, processes, rules, actors, and causal relationships. Responsibility is not determined by emotion; it is formed through an evidence chain, an action chain, and a causality chain.

The AI era must follow the same logic.

If AI is widely used in finance, medicine, law, recruitment, insurance, public services, platform governance, and enterprise management, then after AI makes a mistake, we cannot stop at vague expressions such as “model hallucination,” “system error,” or “algorithmic judgment.” The more AI participates in important affairs, the more it needs a responsibility chain that can be traced, replayed, reviewed, and held accountable.

This is precisely the issue that AI verifiability structures must continue to address.

Previous articles have already proposed that, in the AI era, we need to distinguish among three types of facts: deterministic facts, procedural facts, and cognitive facts. Deterministic facts need to be verified. Procedural facts need to be replayed. Cognitive facts need to be reviewed. This article further explains that these three types of facts do not exist in isolation; together, they form the foundation of the AI responsibility chain.

But fact classification alone is not enough.

If facts are only internal records, they may still be tampered with, selectively disclosed, reconstructed after the event, or disconnected from responsibility. What truly supports an AI responsibility chain is not merely the existence of facts, but the entry of those facts into a verifiable architecture.

The core issue of the AI responsibility chain is not to make AI itself bear responsibility, but to use a verifiable structure to identify the relationships among humans, institutions, systems, rules, and actions.

AI is not a legal person. It has no life, no assets, cannot truly be punished, and cannot bear moral or legal consequences. The real bearers of responsibility remain humans and institutions. The meaning of the AI responsibility chain is not to shift responsibility onto machines, but to prevent humans and institutions from hiding responsibility inside machines.

I. The AI Responsibility Chain Begins with Human Society’s Methods of Accountability

Human society has already developed many methods of accountability. Although different fields use different language, their underlying structures are highly similar.

Legal accountability examines four questions: what the facts are, what the conduct was, whether there is causality between conduct and consequence, and who the responsible subject is.

Audit accountability examines four questions: whether vouchers are authentic, whether accounts are consistent, whether processes are compliant, and who caused or allowed the anomaly.

Medical accident accountability examines four questions: whether the diagnostic basis was sufficient, whether the treatment process was compliant, whether risks were disclosed, and whether the physician and institution fulfilled their duty of care.

Financial risk accountability examines four questions: whether the transaction was real, whether the authorization was valid, whether risks were identified, and whether approval and supervision were in place.

Behind these methods is one common point: accountability must be layered.

The first layer is the foundational fact. Did the event occur? Who participated? When did it happen? Are the data, vouchers, and records authentic?

The second layer is the behavioral process. What did the person or system do? Was it carried out according to rules? Is there a record? Can it be replayed?

The third layer is judgment and decision-making. On what facts was the judgment based? Was the judgment reasonable? Was there obvious neglect, abuse, misleading conduct, or overreach?

The fourth layer is the responsible subject. Who designed the rules? Who provided the data? Who deployed the system? Who authorized the use? Who ultimately adopted the result? Who benefited from it? Who should bear the consequences?

This is the basic logic of human accountability: we do not first look for an abstract object to take the blame. Instead, we trace layer by layer through facts, processes, judgments, and subjects.

The AI responsibility chain should begin from the same point.

II. AI Errors Cannot Be Reduced to “The AI Was Wrong”

After AI makes a mistake, if we only say “the AI was wrong,” responsibility is flattened.

If an AI medical assistance system gives an incorrect recommendation, the problem may not lie in the model itself. The input test data may have been incomplete; medical records may have been missing; the system may have failed to display a risk warning; the doctor may have relied too heavily on AI; the hospital may not have established a review procedure; or the product design may have packaged a probabilistic judgment as a deterministic conclusion.

If an AI financial risk-control system lets an abnormal transaction pass, we cannot simply say “the model was wrong.” Transaction data may have been deliberately split; rule thresholds may have been set too loosely; historical training data may have been biased; system alerts may have been ignored by humans; or management may have lowered risk-control standards for business growth.

If an AI recruitment system rejects a candidate, we cannot simply say “the algorithm discriminated.” We need to examine whether the input data were accurate, whether the screening rules were compliant, whether the model’s judgment basis was explainable, whether humans reviewed the result, and whether the company treated the AI recommendation as the final decision.

These examples show that AI errors are often not single-point errors. They are usually the result of the interaction among facts, procedures, rules, models, human judgment, and organizational responsibility.

Therefore, the AI responsibility chain must answer three foundational questions.

First, did the error occur at the factual layer? This asks whether deterministic facts such as input, data, records, identity, authorization, transactions, vouchers, time, and status were true, complete, and consistent.

Second, did the error occur at the procedural layer? This asks whether the AI executed within the authorized scope, called the correct tools, followed the rules, left process records, and whether there was overreach, omission, anomaly, or a failure to trigger a necessary alert.

Third, did the error occur at the cognitive layer? This asks whether the AI’s analysis, judgment, reasoning, recommendation, scoring, prediction, or conclusion was reasonable; whether it treated opinions as facts; whether it packaged probabilistic judgments as deterministic conclusions; and whether human review was absent.

These three questions correspond precisely to the three types of facts proposed in previous articles: deterministic facts, procedural facts, and cognitive facts.

Once fact classification enters an accountability scenario, it is no longer merely an epistemological classification; it becomes a structure of responsibility.

III. Deterministic Facts: The Starting Point of the Responsibility Chain

Any responsibility trace must begin with deterministic facts.

Deterministic facts are facts that can be verified. For example: whether an identity is authentic, whether authorization is valid, whether a transaction occurred, whether funds arrived, whether a document was submitted, whether a timestamp exists, whether a record was modified, whether an account balance is consistent, and whether a specific operation was initiated by a specific subject.

In human accountability, deterministic facts are equivalent to the evidentiary foundation. Without an evidentiary foundation, responsibility judgments become speculation.

In AI scenarios, deterministic facts are even more important. AI outputs often create the illusion that something has already been processed. The more complex a system becomes, the more easily people overlook whether the foundational facts are reliable.

If AI makes a judgment based on incorrect data, responsibility should not first be placed on the cognitive conclusion. We must ask where the data came from, who provided it, who reviewed it, and who allowed it to enter the system.

If AI executes a task based on outdated rules, responsibility should not be placed only on the model. We must ask who maintains the rule base, whether the version was updated, and whether the system warned that the rule had become invalid.

If AI automatically completes a financial operation, responsibility must first verify whether authorization existed, whether permissions matched, whether the amount exceeded limits, whether the account was correct, and whether the execution time can be checked.

Therefore, the first layer of the AI responsibility chain is fact verification.

This layer answers the following question: are the foundational facts on which AI relies true, complete, tamper-resistant, and independently reviewable by a third party?

In verifiable finance, this layer can be strengthened through ledger records, hashes, citation chains, a public credit root, and consistency between on-chain and off-chain records. In general AI scenarios, it at least requires timestamps, logs, version records, data sources, authorization records, and auditable credentials.

Without verification of deterministic facts, later process replay and cognitive review lack a foundation.

IV. Procedural Facts: The Trunk of the Responsibility Chain

Deterministic facts tell us what happened. But responsibility tracing must also know how the event happened.

This is the role of procedural facts.

A procedural fact is not a single result. It consists of the steps, calls, paths, permissions, rules, anomalies, pauses, human interventions, and system feedback involved in task execution.

Human society has long valued process records. Audits examine procedures. Medicine examines medical records and operation records. Courts examine whether procedures are lawful. Enterprise management examines approval chains. Traffic accident investigations examine driving records, surveillance footage, and scene trajectories.

In the AI era, process records are even more important, because an AI output by itself often cannot explain how the conclusion was reached.

If an AI assistant gives incorrect legal advice, we cannot look only at the final sentence. We must also examine which materials it called, which conditions it ignored, whether it cited outdated laws, and whether it incorrectly applied a similar case to the current issue.

If an AI trading system performs an erroneous operation, we cannot look only at the final loss. We must examine what instruction it received, which risk-control judgments it passed through, whether approval was triggered, whether human confirmation occurred, and whether it exceeded its authorization boundary.

If an AI customer-service system wrongly rejects a user appeal, we cannot look only at the rejection result. We must examine what materials the user submitted, how the system classified the case, how the rules matched, whether human review was available, and whether a batch misjudgment occurred.

Therefore, the second layer of the AI responsibility chain is process replay.

Process replay is not meant to record a pile of running logs. Its purpose is to answer whether AI acted within the correct authorization, the correct rules, the correct scope, and the correct process.

This layer should at least record who initiated the task, what the authorization scope was, what input materials were used, which tools or models were called, which rule version was used, what intermediate steps occurred, whether anomalies appeared, whether human review was triggered, how the final output was formed, and who adopted or executed the output afterward.

If the process is invisible, responsibility breaks. If the process cannot be replayed, the aftermath becomes mere dispute. If the process leaves no trace, institutions can push responsibility onto the model, while the model cannot truly bear responsibility.

Therefore, the core of verifiable AI execution is not merely to make AI safer, but to allow the AI execution process to enter a structure of responsibility.

V. Cognitive Facts: The Most Complex Layer of the Responsibility Chain

The most complex aspect of AI is not that it executes a clear instruction, but that it participates in judgment.

Summarizing, analyzing, comparing, recommending, scoring, diagnosing, predicting, assessing risk, evaluating credit, screening candidates, offering legal advice, and providing investment advice are all cognitive activities.

Cognitive activities differ from deterministic facts. Whether an opinion is correct often cannot be verified in the same way as an account balance. Whether a judgment is reasonable also cannot be proven entirely by a hash or timestamp.

But this does not mean cognitive judgments cannot enter the responsibility chain.

The key to cognitive facts is not to treat the opinion itself as a deterministic fact, but to treat “how a cognitive conclusion was formed” as the object of review.

For example, if AI gives a medical risk judgment, responsibility tracing does not necessarily require proving that the judgment is absolutely correct. Instead, it examines which test results it relied on, what rules it applied, whether it indicated uncertainty, whether it recommended physician review, and whether it concealed risk boundaries.

If AI gives a credit score, responsibility tracing should not only ask whether the score was high or low. It should examine whether the data were compliant, whether the variables were reasonable, whether discriminatory features existed, whether the user had a right to appeal, and whether there was final human judgment.

If AI provides a legal opinion, responsibility tracing should not only ask whether the conclusion matches a court judgment. It should examine whether the cited materials were real, whether the laws were current, whether the reasoning jumped, whether possibilities were stated as certainties, and whether the need for professional lawyer review was disclosed.

This is cognitive review.

Cognitive review does not require AI conclusions to always be correct. It requires AI judgments to have sources, boundaries, reasons, uncertainty warnings, and, in major scenarios, review by humans or external verification AI.

Here a new question arises: if another AI is used to review the first AI, who reviews that reviewing AI?

This problem cannot be solved through the infinite recursion of “finding another AI.” The real solution is to establish a human-in-the-loop structure and an objection mechanism. External verification AI can help discover factual errors, procedural anomalies, reasoning gaps, and missing risk warnings, but the final right of verification should remain with humans and responsible institutions. Especially in finance, medicine, law, recruitment, insurance, public services, and similar scenarios, an AI review conclusion must not automatically become the final responsibility judgment.

In other words, external verification AI may participate in the responsibility chain, but it cannot replace the responsible subject.

Within the AI responsibility chain, cognitive facts are the easiest zone for responsibility to escape. Many institutions can say, “This was only an algorithmic recommendation.” But if that recommendation in fact affects loans, medical treatment, recruitment, insurance, penalties, asset disposal, or public services, then it cannot remain merely a “recommendation.”

As long as an AI judgment enters real-world consequences, it must enter the responsibility chain.

VI. The Responsibility Chain Is Not Linear, but a Multi-Subject Structure

The AI responsibility chain is not a simple line traced from AI back to a single person. It is often a multi-subject responsibility structure.

It includes at least the following types of subjects:

• Data providers: If foundational data are wrong, missing, polluted, or used without authorization, the data provider or data manager may bear responsibility.
• Model developers: If the model has known defects, false claims, undisclosed major limitations, or insufficient safety design, the developer may bear responsibility.
• System deployers: If an institution uses AI in an unsuitable scenario, or fails to establish necessary supervision, review, and permission controls, the deployer may bear responsibility.
• Rule-makers: If an enterprise or institution creates unreasonable rules and causes AI to optimize toward the wrong objective, responsibility cannot be shifted onto AI.
• Operators: If a user enters an incorrect instruction, ignores warnings, acts beyond authority, or deliberately abuses AI, the operator should bear responsibility.
• Reviewers: If the system requires human review but the reviewer fails to perform that duty, the reviewer is also within the responsibility chain.
• Decision-makers: If AI merely gives a recommendation, but humans or institutions ultimately adopt it and produce consequences, the final decision-maker cannot shift all responsibility onto AI.
• Beneficiaries or managers: If an institution uses AI to reduce costs, expand business, or increase profits, it should also bear corresponding governance responsibility. It cannot enjoy efficiency while escaping consequences.

This shows that the AI responsibility chain is not meant to find a scapegoat, but to clarify the boundaries of responsibility among different subjects at different links.

The more AI enters society’s core systems, the less responsibility can be allowed to remain in a state of mutual deflection among “model provider,” “platform,” and “user.”

A mature AI system must allow responsibility to land layer by layer along facts, processes, judgments, authorization, and consequences.

VII. The Responsibility Chain Must Be Built on a Verifiable Architecture

Without a verification chain, there is no responsibility chain.

This is the most important judgment of this article.

Traditional society can hold people accountable because it has contracts, ledgers, vouchers, medical records, surveillance footage, approval records, audit reports, witness testimony, and legal procedures.

In the AI era, if there is no corresponding verification structure, responsibility will be swallowed by the black box.

But this point must be further clarified: a verification chain is not simply a pile of logs, nor is it enough for an institution to keep an internal record. Internal logs, internal approvals, and internal audits all have value. But if key facts can only be stored, interpreted, and selectively disclosed by the institution itself, the responsibility chain remains incomplete.

A real AI responsibility chain must be built on a verifiable architecture.

A verifiable architecture should include at least four principles.

First, rule parity. The rules governing AI authorization, execution, review, penalty, appeal, and re-examination must not become a unilateral black box. Different participants should at least know the basic rules, scope of application, and available remedies. Otherwise, AI may become a form of automated power that cannot be questioned.

Second, separation of powers. Rule-making, task execution, and verification oversight must not be fully monopolized by the same party. Those who make the rules should not fully execute them by themselves, and those who execute should not fully prove themselves by themselves. The more important an AI system is, the more necessary it becomes to separate rules, execution, and verification to an appropriate degree.

Third, whole-system verifiability. Verification cannot look only at results. It must also verify inputs, authorization, processes, rules, anomalies, review, and responsibility. Looking only at the final AI output cannot form a responsibility chain; looking only at internal logs is also insufficient to prevent ex post modification and selective interpretation. A real responsibility chain must cover the full process from task initiation to result adoption.

Fourth, consistency between on-chain and off-chain records. Here, “on-chain and off-chain” is not limited to the blockchain sense. More accurately, it refers to consistency between real-world facts and verifiable records: what happened in the real world, what the system recorded, and what the external verification structure anchored must correspond to one another. Otherwise, off-chain responsibility can be severed, and on-chain records will lose their meaning.

These four principles show that the AI responsibility chain is not merely a compliance document, but an institutional architecture.

Therefore, the AI responsibility chain requires at least four categories of records.

First, factual records. These record input data, authorization information, identity information, timestamps, file versions, transaction status, and sources of foundational facts.

Second, process records. These record task steps, model calls, tool calls, rule versions, intermediate outputs, anomalies, and human intervention.

Third, cognitive records. These record the basis of AI judgments, cited materials, uncertainty warnings, opposing views, review opinions, and final human judgment.

Fourth, authorization and responsibility records. These record who initiated the task, who approved execution, who reviewed the result, who adopted the recommendation, and who bears final decision-making responsibility.

Authorization and responsibility records are not another independent category of “fact classification.” They are records of authorization, approval, adoption, review, and final decision-making. These actions themselves are still facts that can be recorded and verified, but they have special significance within the responsibility chain: they determine where responsibility ultimately lands.

If hashes, timestamps, tamper-resistant records, external verification AI, third-party audits, and a public credit root are further introduced, the responsibility chain is no longer merely an internal record; it can become an externally verifiable structure.

This is the key difference between “internal compliance” and “external verifiability.”

Internal compliance can only show that an institution says it has a process.

External verifiability means that key facts, key processes, and key responsibilities can be independently reviewed.

In the AI era, a truly reliable responsibility structure cannot rely only on institutional self-certification. It must move toward a verifiable responsibility chain.

VIII. The Goal of the AI Responsibility Chain Is Not to Restrict AI, but to Make AI Safe to Use

Some may believe that emphasizing the responsibility chain will restrict AI development.

This view is shortsighted.

Without a responsibility chain, AI can spread quickly in low-risk scenarios, but in high-risk scenarios it will inevitably encounter institutional resistance. Finance, medicine, law, insurance, public services, enterprise management, asset disposal, and similar fields cannot accept an unaccountable black-box system over the long term.

The responsibility chain is not an obstacle to AI development. It is the passport for AI to enter key fields.

An AI that cannot be held accountable can only be an auxiliary tool.

An AI that can be replayed, reviewed, audited, and held accountable can enter finance, medicine, law, and public governance.

This logic is consistent with verifiable finance.

Traditional finance operates through institutional credit, but the cost of trust is extremely high. Verifiable finance does not abolish financial institutions; it makes key financial facts verifiable, thereby reducing trust costs and increasing institutional transparency.

The AI responsibility chain works in the same way. It does not abolish AI, nor does it lock AI away. It allows key AI actions to enter a verifiable structure, so that humans and institutions dare to use it, regulators can supervise it, users can appeal, and responsibility can land.

Without a responsibility chain, the stronger AI becomes, the more uneasy society becomes.

With a responsibility chain, the stronger AI becomes, the more capable institutions become of absorbing its power.

IX. From Fact Classification to Responsibility Chain: A Key Advance in AI Verifiability Theory

Previous articles proposed fact classification in order to explain that different facts cannot be verified by the same method.

This article further explains that fact classification is not only a question of verification method, but also a question of responsibility structure.

Deterministic facts determine whether the responsibility chain has a foundation.

Procedural facts determine whether the responsibility chain can be replayed.

Cognitive facts determine whether the responsibility chain can be reviewed.

Authorization and responsibility records determine where responsibility ultimately lands.

The verifiable architecture determines whether the responsibility chain can be externally confirmed.

Therefore, the AI responsibility chain can be summarized in one sentence:

Start with verification of foundational facts, replay the execution process, review cognitive judgments, finally assign responsibility to humans and institutions, and use a verifiable architecture to make the entire responsibility chain externally confirmable.

This also shows that the greatest danger in the AI era is not that AI will make mistakes. Human society has long known that any system can make mistakes. The real danger is that after AI makes a mistake, no one can clearly explain why it was wrong, where it was wrong, who should be responsible, and how to prevent it from happening again.

The verifiable structure is meant to solve precisely this problem.

AI should not become a black hole of responsibility.

AI should become an execution and cognition system that can be recorded, replayed, reviewed, and held accountable.

Human society accepts technological innovation not because technology never makes mistakes, but because mistakes can be discovered, corrected, attributed, and absorbed by institutions.

AI must also enter this logic.

If the previous stage of AI competition was mainly a competition of capability, the next stage of AI competition will inevitably become a competition of responsibility. Whoever can make AI more verifiable, more replayable, more auditable, and more accountable will be able to bring AI into truly important institutional scenarios.

This also opens the direction for the next discussion: since different industries have different responsibility chains, AI in different industries should not be classified only by model capability. Its design should be reverse-engineered from the verifiable structure.

Financial AI needs to form a responsibility chain around authorization, transactions, accounts, risks, and audits.

Medical AI needs to form a responsibility chain around medical records, examinations, diagnoses, treatment recommendations, and physician review.

Legal AI needs to form a responsibility chain around facts, evidence, statutes, reasoning, and lawyer responsibility.

Recruitment AI needs to form a responsibility chain around data sources, screening rules, scoring bases, appeal mechanisms, and final decisions.

Public-service AI needs to form a responsibility chain around rule disclosure, procedural due process, human review, and channels for remedy.

Therefore, AI in different industries should not only be asked what it can do. It should also be asked how it can be verified, how it can be held accountable, and where its responsibility chain may break.

This is the meaning of the AI responsibility chain.

It is not a compliance slogan. It is the structure that AI must complete before it can move from a tool into institutional infrastructure.

Without a responsibility chain, AI can only provide answers.

With a responsibility chain, AI can enter the world of responsibility.

Chapter 15 Verifiable Thinking Will Lead AI Toward a Diverse Future

— A Theoretical Reflection on the AI Giant Narrative

AI is becoming the most powerful technological variable of our time.

It can write, program, translate, search, summarize, reason, invoke tools, generate plans, and even begin to enter enterprise management, financial analysis, medical assistance, legal services, educational training, and public governance. The emergence of large models has made many people feel, perhaps for the first time, that software is no longer merely software, tools are no longer merely tools, and machines seem to be acquiring a general intelligence capable of moving horizontally across different domains.

Precisely because of this, a powerful ‘giant narrative’ has gradually formed in the AI industry. One voice argues that AI will continue to become larger, stronger, and more general, eventually becoming the unified intelligent entrance for all software, all industries, and all knowledge work. Another voice worries that AI will ultimately become a superintelligence, break free from human control, and threaten humanity itself.

These two voices appear to be opposed, but they share the same underlying premise: the future will produce an increasingly large, increasingly general, and increasingly centralized AI giant. The optimists believe this giant will bring prosperity; the doomsayers believe it will bring catastrophe.

But verifiable thinking offers a third judgment: a high-responsibility society will not allow an unverifiable AI giant to become the basic structure of society. AI can be powerful, but it cannot be unverifiable. AI can be general, but it cannot be without boundaries. AI can become infrastructure, but it cannot become a sovereign force that is beyond review, beyond constraint, and beyond responsibility.

This essay does not ask whether AI will continue to develop, nor whether large models have value. It asks a more fundamental question: after AI enters real society, how will institutions, responsibility, and verification structures shape the form of AI?

My judgment is this: verifiable thinking will make AI diverse.

This is not a futurist prediction. It is a form of reasoning from institutional structure.

I. The Position of This Essay Within the Series on Verifiable Thinking

This is not an isolated commentary on AI. It is a further development of the verifiable-thinking series on AI.

In previous essays, I have discussed the relationship between AI and verification from different angles.

AI Must Also Be Verified: From Verifiable Finance to Verifiable Execution put forward a basic judgment: once AI moves from answering questions to executing tasks, we cannot look only at the output. We must record authorization, process, basis, and responsibility.

How to Build a Verifiable Structure for AI further discussed how AI execution can leave process records, form replay paths, and allow external parties to review key AI actions.

Humans Need the Right to Verify AI emphasized that when AI participates in finance, medicine, law, recruitment, platform penalties, public services, and other high-responsibility scenarios, people cannot merely passively accept AI outputs. They must retain the final right to review and appeal.

Verifiable Thinking Is the Dragon-Locking Chain of the AI Era explained at a more macro level that AI is a powerful force, but a powerful AI without verification structures may become a source of black-box execution, responsibility escape, power expansion, and cognitive enslavement. Verifiable thinking is not anti-AI. It is the institutional rein that allows AI to be guided, reviewed, and held accountable.

These essays address different questions along the same line: Why must AI be verified? How can AI execution be verified? Why do humans need verification rights? Why is verifiable thinking an institutional constraint in the AI era?

This essay advances the discussion further and asks another more basic question: if AI must be constrained by verifiable structures, what form will it ultimately take?

Therefore, this essay does not repeat the proposition that ‘AI must be verified.’ Instead, it starts from that premise and repositions the future form of AI. If AI must be verified, and if verification structures differ across domains, then AI will not evolve only in the direction of a larger and larger giant. It will be shaped into a diverse ecosystem by different institutions, different responsibilities, and different verification structures.

For this reason, this essay is a positioning essay on AI. It starts from the institutional constraints AI faces after entering society and derives why AI will become diverse.

AI has general capability, so the giant narrative arises. AI enters a high-responsibility society, so it must be verified. Different institutions generate different responsibilities, so verification structures differ. Different verification structures produce diverse AI forms. The general capability of large models remains, but it returns to the position of infrastructure.

This is the complete logic of the essay.

II. The AI Giant Narrative Comes From General Capability

The reason large models have triggered the imagination of an AI giant is not without foundation.

Software in the past was functional. Accounting software handled accounts; medical systems managed medical records; legal databases retrieved cases; office software processed documents. The boundaries among different software systems were clear. Users had to understand processes, enter information, click buttons, and judge results.

Large models have changed this relationship. They can understand tasks in natural language, call tools, generate code, connect databases, operate software, and decompose complex tasks into steps. It is therefore natural for people to imagine that if one model can understand all tasks, call all software, master all knowledge, and coordinate all processes, might it become a unified intelligent layer above all software?

This is the technical source of the AI giant narrative.

From the perspective of capability, this imagination is not absurd. AI does have strong generality. Language, code, logic, text, images, tool use, and knowledge integration all have cross-domain characteristics. The stronger a large model becomes, the more it resembles a general capability base that can enter different domains.

If one reasons only from capability, it is easy to conclude that AI will become larger, more unified, and more giant-like.

But the problem is that human society is not composed of capability alone. It is also composed of institutions, responsibility, rights, risk, evidence, procedure, and accountability.

What AI can do is a technical question. What AI is allowed to do is an institutional question. How AI actions can be reviewed and how responsibility is assigned after errors occur is a verification question.

The flaw of the giant narrative lies here: it over-derives social form from technical capability, while underestimating the reverse shaping force that institutions and verification structures exert on AI.

III. This Essay Does Not Predict the Future; It Derives the Future

Discussions of AI’s future easily fall into the style of futurism. Futurism often asks: What will appear in the future? Which technologies will become popular? Which trends will expand? Which professions will disappear? Which industries will be restructured?

This style has value, but it also has an obvious problem. Many future judgments appear decades later to have ‘gotten it right,’ but people often remember the correct fragments and forget the incorrect ones. If someone produces enough future imaginings, some fragments will always be selected by history.

A truly powerful theory does not merely guess the future. It explains why a certain future is more likely to be jointly pushed forward by institutions, technology, and social structure.

When Schumpeter discussed capitalism, he did not rely on guessing future industries. He grasped internal mechanisms such as innovation, entrepreneurship, and creative destruction. When Zhang Weiying discusses markets and entrepreneurs, he does not rely on trend intuition alone. He reasons from property rights, incentives, dispersed knowledge, institutional constraints, and entrepreneurial judgment. This essay adopts a similar method.

This essay is not saying, ‘I predict that AI will become diverse.’ It is saying that once AI enters a high-responsibility society, it must be verified; different institutions generate different responsibilities; different responsibilities generate different verification structures; and different verification structures inevitably shape different AI forms.

This is mechanism-based reasoning, not prophecy.

Institutions determine responsibility, responsibility determines verification structures, and verification structures determine AI forms.

Or, more directly: technology determines what AI can do; institutions determine what AI is allowed to do; verification structures determine what AI ultimately becomes.

IV. Adjacent Judgments in Existing Economic Research and the Original Position of This Essay

A theoretical judgment cannot rely only on self-confirmation. If a new proposition is worth putting forward, it is usually not because no one has touched it at all, but because many front-line researchers have already seen partial phenomena without yet forming a unified framework. Therefore, when discussing the future form of AI, we must first conduct a broad intellectual and economic scan.

Existing economic research has already weakened the imagination of a single AI giant from several angles. Task economics decomposes AI’s impact into automation, complementarity, and the creation of new tasks in different settings, rather than treating AI as one unified substitute. Brynjolfsson’s discussion of the ‘Turing Trap’ reminds us that the direction of AI is not limited to imitating and replacing human beings; AI can also be designed to augment human capabilities. The analysis of pro-worker AI by Acemoglu, Autor, Johnson, and others similarly emphasizes that AI should not only replace labor, but can also expand workers’ skills, judgment, and professional capability. Research on AI knowledge aggregation further shows that in multi-topic knowledge environments, a single global aggregator is not necessarily superior to specialized or local aggregators. Research on foundation-model market concentration also shows that although frontier models may tend toward concentration, non-frontier models, downstream applications, and industry-level deployment will still produce differentiated structures.

All of these studies are important. They show that AI’s future does not have only one path in which a larger and larger model absorbs everything. Differences in tasks, labor relations, knowledge aggregation, market structures, and the relationship between upstream foundation models and downstream applications will all create pressures toward AI diversity.

However, these adjacent judgments still mainly proceed from tasks, labor, knowledge, markets, and competition structures. The original position of this essay is not to repeat that ‘AI will have many applications’ or that ‘specialized models will exist.’ It is to reason from verifiable thinking and propose another path of derivation: after AI enters a high-responsibility society, it must face institutions, responsibility, and verification structures. Since verification objects, verification subjects, verification standards, and responsibility boundaries differ across domains, the actual form of AI will necessarily differ as well.

In other words, economic research has already identified several market and knowledge foundations for AI diversity. This essay further points out that AI diversity has a deeper institutional foundation: verifiable structures. Tasks make AI divide labor; markets make AI compete; knowledge structures make AI stratified; but verification structures institutionally shape AI. This is the difference between this essay and ordinary market judgments, professional-model judgments, and existing economic discussions.

V. General Capability Does Not Mean General Responsibility

AI capability can be general, but responsibility cannot be general. This is the key to understanding the future form of AI.

A large model can simultaneously understand finance, medicine, law, auditing, education, and enterprise management. It can read financial statements and summarize medical records; retrieve legal provisions and analyze contracts; write programs and generate market reports.

But understanding the knowledge of a field is not the same as being able to carry the responsibility of that field.

Finance is not merely numerical calculation. It involves accounts, balances, reserves, transactions, authorization, clearing, accounting, auditing, and legal responsibility.

Medicine is not medical knowledge Q&A. It involves medical records, diagnostic bases, treatment plans, risk disclosure, patient consent, physician responsibility, and medical norms.

Law is not simple text matching of statutes. It involves fact-finding, rules of evidence, procedural justice, interpretive boundaries, agency responsibility, and adjudicative consequences.

Auditing is not report summarization. It involves vouchers, contracts, ledgers, revenue recognition, cost allocation, anomaly identification, and responsibility signatures.

Public services are not mere process automation. They involve the source of authority, administrative procedure, disclosure obligations, appeal mechanisms, chains of responsibility, and public supervision.

These fields are not merely collections of knowledge. They are institutional structures. Large models can learn industry knowledge, call industry software, and assist industry professionals, but they cannot automatically absorb industry responsibility simply because they ‘know a lot.’

Knowledge can be absorbed by a model; responsibility cannot. Software can be called by a model; verification cannot be swallowed by a model. Interfaces can be unified; responsibility boundaries cannot.

The giant narrative often mixes these things together. It sees that a large model can call software and assumes that the model can absorb software. It sees that a large model can handle industry tasks and assumes that the model can absorb industries. It sees that a large model can produce explanations and assumes that the model can absorb verification.

But this is wrong. AI can become a center of capability, but it cannot become the center of all responsibility. AI can become a unified entrance, but it cannot become the final judge of verification for all industries.

VI. Verifiable Structures Are the Channels Through Which AI Enters Real Society

Water has no fixed shape. It can flow, spread, converge, and overflow boundaries.

But once water enters the real world, it is constrained by river channels, embankments, terrain, and engineering structures. A river channel is not the enemy of water. It is the condition that allows water to be used, guided, controlled, and distributed.

The general capability of AI is similar.

In an imagination without institutional constraints, AI can be described as an omnipotent giant. It is like water without boundaries, seemingly capable of flowing into all domains, absorbing all software, and completing all tasks.

But once AI enters finance, medicine, law, auditing, insurance, recruitment, education, public services, and other real domains, it must enter different institutional channels.

Financial institutions require AI to leave records of accounts, transactions, authorization, reserves, and accounting. Medical institutions require it to leave diagnostic bases, medical-record evidence, risk warnings, and professional responsibility trails. Legal institutions require it to distinguish facts, evidence, procedure, rules, and interpretive boundaries. Auditing institutions require it to preserve vouchers, ledgers, contracts, anomalies, and review paths. Public service institutions require it to accept authority constraints, procedural review, appeal mechanisms, and responsibility tracking.

This is the verifiable structure.

A verifiable structure is not a simple technical plug-in, nor is it merely adding an explanation button to AI. It is the institutional channel that AI must accept after entering real society.

Without verification structures, AI can be imagined as an unbounded giant. With verification structures, AI can become a tool that is reviewable, accountable, and capable of entering a high-responsibility society.

Therefore, the future of AI is not determined solely by model capability. It is determined by AI’s institutional position after it enters society.

VII. Use Determines Software Differentiation; Institutions Determine Verification Differentiation

AI applications in different fields will first differentiate because their uses differ.

Financial AI, medical AI, legal AI, educational AI, enterprise-management AI, and auditing AI naturally have different functions. Programmers and product teams can upgrade existing software systems, connect AI to existing workflows, and make software more intelligent, more automated, and easier to use.

This is application-layer differentiation.

But verifiable thinking emphasizes not only use-based differentiation, but verification-based differentiation. Use-based differentiation answers the question: what is AI used for? Verification-based differentiation answers the question: why should AI be trusted, how can it be reviewed, and who is responsible when something goes wrong?

Ordinary industry-software upgrades can be completed by programmers according to requirements. But the verification conditions for high-responsibility AI cannot be fully designed by ordinary programmers sitting in an office. Verification conditions are institutional questions.

The verification conditions for financial AI must be jointly defined by financial experts, banks, auditors, regulators, and legal responsibility bearers. The verification conditions for medical AI must be jointly defined by doctors, hospitals, medical norms, patient rights, and medical responsibility systems. The verification conditions for legal AI must be jointly defined by lawyers, judges, evidentiary rules, procedural institutions, and judicial responsibility. The verification conditions for auditing AI must be jointly defined by accounting standards, auditing standards, voucher systems, enterprise financial practice, and legal responsibility. The verification conditions for public-service AI must be jointly defined by administrative law, public procedure, appeal mechanisms, supervisory systems, and responsible agencies.

Therefore, AI diversity is not merely the result of market competition, nor is it merely caused by developer preference. It comes from a deeper structure: use determines software differentiation, institutions determine verification differentiation, and verification differentiation ultimately determines the diversity of AI forms.

VIII. A Map of Verification Structures: Why AI Will Necessarily Flourish in Many Forms

If there were no verification structures, people could imagine one giant general model that handles everything. But once verification is introduced, a completely different picture emerges.

Table 1: Verification Structure Map Across Different Industries

Industry	Main Objects of Verification	Main Verification Subjects	Direction in Which AI Forms Are Shaped
Finance	Accounts, balances, reserves, transactions, authorization, accounting, audit replay	Banks, audit institutions, regulators, transaction-related parties	Financial execution AI, risk-control AI, audit-verification AI, reserve-verification AI
Medicine	Medical records, diagnostic bases, treatment plans, risk disclosure, patient consent	Doctors, hospitals, patients, medical-norm setters, regulators	Medical-assistance AI, diagnostic-review AI, risk-warning AI
Law	Facts, evidence, procedure, applicable rules, interpretive boundaries, agency responsibility	Lawyers, judges, parties, judicial institutions	Legal-retrieval AI, evidence-review AI, procedure-review AI
Auditing	Vouchers, contracts, ledgers, revenue recognition, anomaly records, responsibility signatures	Accountants, audit institutions, enterprises, regulators	Ledger-inspection AI, anomaly-detection AI, audit-replay AI
Public Services	Source of authority, administrative procedure, public disclosure, appeal, chain of responsibility	Administrative agencies, the public, supervisory bodies, courts	Government-service AI, procedural-compliance AI, appeal-review AI

Here, ‘verification subjects’ mainly refers to the combination of responsible review subjects and technical verification executors. AI can participate in inspection, warning, and recording, but final legal responsibility should still be borne by people, institutions, or corresponding responsible subjects.

This map shows that AI diversity is not derived from model competition alone, but from verification structures. Different domains require different verification structures, so AI will necessarily flourish in many forms.

Some AI systems will lean toward execution, while others will lean toward judgment. Some will lean toward auditing, while others will lean toward recording. Some will lean toward review, while others will lean toward risk control. Some will lean toward professional assistance, while others will lean toward public supervision.

The future will not be one AI giant handling everything, but a verifiable AI ecosystem jointly composed of general models, specialized models, verification models, rule systems, audit structures, and responsible subjects.

IX. Why This Is Different From the Ordinary Specialized-Model Argument

Some may say that specialized models, industry models, and vertical models are already common judgments in the AI industry, and that there is nothing new here.

This objection sees only the surface. The ordinary specialized-model argument usually starts from engineering efficiency: general models are too large, costs are too high, industry data are insufficient, professional scenarios are complex, so vertical models are needed. This judgment is not wrong, but its level is insufficient.

This essay is not discussing engineering division of labor, but institutional shaping. Engineering division solves efficiency problems; verifiable structures solve credibility problems.

As the following table shows, the ordinary specialized-model argument and the verifiable-thinking argument differ in their starting points and core questions.

Table 2: Differences Between the Ordinary Specialized-Model Argument and the Verifiable-Thinking Argument

Dimension	Ordinary Specialized-Model Argument	Judgment From Verifiable Thinking
Starting point	Cost, efficiency, data, performance	Institutions, responsibility, verification, accountability
Core question	How can a model perform tasks better?	Why should AI be trusted, and how can it be reviewed?
Reason for differentiation	Different industry tasks and different data	Different verification objects, verification subjects, verification standards, and responsibility boundaries
Final form	Multiple vertical applications or vertical models	An ecosystem composed of general models, specialized models, verification models, rule systems, and audit structures

This distinction shows that the essay is not discussing the ordinary judgment that ‘there will be many industry applications.’ It is discussing how a high-responsibility society reshapes the actual form of AI through institutions and verification structures.

Specialized models can emerge because of efficiency, data, or cost. But the real differentiation of high-responsibility AI comes from institutions, responsibility, and verification.

If a medical AI merely answers medical questions, it can be an ordinary specialized model. But if it participates in diagnostic recommendations, it must enter the medical responsibility structure.

If a financial AI merely summarizes market news, it can be an ordinary industry model. But if it participates in lending, risk control, trading, reserves, or auditing, it must enter the financial verification structure.

If a legal AI merely retrieves statutes, it can be an ordinary legal tool. But if it generates legal opinions, participates in litigation strategy, or affects rights and obligations, it must enter the legal responsibility structure.

Therefore, the AI diversity discussed in this essay does not mean merely that ‘there will be many industry applications.’ It means that after AI enters a high-responsibility society, it will be shaped into different forms by different institutions, different responsibilities, and different verification structures.

X. The AI Giant Narrative Ignores That Responsibility Cannot Be Concentrated Without Limit

The greatest blind spot of the AI giant narrative is that it mistakes concentration of capability for concentration of responsibility.

Technically, the stronger a large model becomes, the more tasks it seems capable of assuming. It can unify interaction entrances, tool calls, task decomposition, knowledge retrieval, and execution workflows.

But institutionally, the higher the responsibility required by a task, the less society can allow responsibility to be concentrated without limit in an unverifiable model.

If an AI is responsible for medical advice, it must be able to explain its basis, risks, and responsibility boundaries. If an AI is responsible for loan approval, it must be able to explain data sources, scoring rules, authorization scope, and appeal paths. If an AI is responsible for legal opinion, it must be able to explain the factual basis, source of evidence, applicable rules, and interpretive boundaries. If an AI is responsible for public services, it must be able to explain the source of authority, whether procedure was compliant, and whether the result can be appealed.

This cannot be automatically solved simply because a large model is ‘smarter.’ The issue is not how intelligent AI is, but whether society will allow an unverifiable intelligent system to directly carry high-responsibility power.

If an unverifiable giant AI simultaneously acts as executor, explainer, verifier, and judge, it will create a new black-box power: it makes the decision, explains the decision, reviews the decision, and judges whether it complied with rules.

That is not a verifiable society. It is a black-box empire.

Verifiable thinking cannot accept such a structure. AI can participate in execution and in verification, but it cannot monopolize verification, and it cannot become the final judge of its own behavior. Genuine verification must allow external subjects outside the model to review facts, processes, authority, rules, and responsibility.

Otherwise, so-called ‘AI self-verification’ is merely black-box self-certification. It replaces ‘trusting institutions’ with ‘trusting models.’

XI. Large Models Can Become Entrances, But an Entrance Is Not a Sovereign

Verifiable thinking does not deny that large models will continue to develop. On the contrary, large models are likely to become important infrastructure in the future AI ecosystem.

They can undertake language understanding, cross-domain communication, task decomposition, code generation, tool use, knowledge integration, and general reasoning. They may also become the unified entrance for many software systems and services.

But a unified entrance is not unified responsibility. Unified scheduling is not unified verification. Unified capability is not a unified institution.

This is similar to the internet. The internet connects everything, but it has not replaced all institutions. It supports e-commerce, finance, media, social networking, government services, education, cloud computing, and many industrial applications, but different applications remain subject to different laws, regulations, and industry rules.

The internet became infrastructure, not a sovereign that rules humanity. Large models will likely be similar.

Of course, the internet has also produced large platforms, but platforms did not thereby eliminate institutional responsibilities in different fields. Likewise, large models may become powerful capability platforms, but they cannot eliminate the verification structures of finance, medicine, law, auditing, and public services.

Their general capability will become a foundational tool that supports various industry AI systems, specialized models, verification models, audit systems, and intelligent applications. But what truly determines how AI enters society is not the model itself, but institutions, responsibility, and verification structures.

Large models provide general capability; industry models handle specialized tasks; verification models inspect processes and bases; rule systems define boundaries of authority; audit systems preserve chains of evidence; and humans and institutions retain final responsibility.

This does not deny large models. It repositions them. Large models will not disappear and will not fail, but they should not be imagined as giants that rule all industries, absorb all responsibility, and replace all institutions. They are more like the internet: infrastructure supporting a diverse AI ecosystem.

XII. Verifiability Does Not Mean Unlimited Disclosure or the Same Intensity of Verification Everywhere

When discussing verification, two questions must be clarified: Does verifiable AI require complete openness of models? Does it require all scenarios to use verification of the same intensity?

The answer to both questions is no.

Verifiability does not mean unlimited disclosure. Verification is higher than disclosure, and higher than open source. AI companies can retain model parameters, training data, trade secrets, and technical details. Not everything needs to be disclosed to the whole society.

But high-responsibility AI behavior cannot hide inside a black box. Any AI behavior that affects rights, property, responsibility, opportunity, or public interest must be disclosed to the extent necessary to verify responsibility.

At minimum, this should include: what decision the AI made; which facts it relied on; what rules it used; who authorized it to act in that way; which tools it called; whether it exceeded authority; whether the process can be replayed; whether the result can be appealed; and who bears responsibility for errors.

At the same time, verifiability does not require all AI behavior to enter verification structures of the same intensity. An AI that writes poems, creates images, or organizes ordinary materials does not require a complex verification structure. An AI that decides loans, medical advice, legal opinions, insurance claims, platform penalties, or public-service eligibility must have reviewable bases and responsibility chains.

Low-responsibility scenarios can use light verification; high-responsibility scenarios must use strong verification. Verification is not meant to destroy efficiency. In high-responsibility scenarios, it uses controllable cost to replace irreversible risk. AI does not need unlimited openness, but it must be open enough to verify responsibility.

Models can retain trade secrets, but behaviors, bases, authority, processes, and responsibility must enter verifiable structures. AI that truly enters a high-responsibility society cannot rely only on company promises, model self-explanations, or blind user trust. It must accept layered disclosure, process replay, external review, and responsibility tracking.

XIII. The Fact That Verifiable AI Has Not Yet Been Fully Realized Does Not Mean the Direction Is Wrong

Some may object that verifiable AI has not yet been fully realized, and therefore it cannot be used to reflect on the large-model route.

This objection appears forceful, but it does not stand. This essay does not claim that verifiable AI is already a mature industry. It discusses the structural requirements that AI must face after entering a high-responsibility society.

If ‘not yet fully realized’ were a reason a subject could not be theoretically discussed, then strong AI, superintelligence, AI restructuring all industries, AI saving the world, and AI destroying the world would likewise be unqualified for discussion. The AI giant narrative itself is a future-oriented judgment. Since the large-model route can discuss futures of strong AI that have not yet been realized, verifiable thinking can certainly discuss verification structures that have not yet matured.

More importantly, verifiable AI is not a fantasy. Many technical components already exist: logging, permission control, version management, tool-call records, audit trails, process replay, external review, hash-based evidence preservation, and responsibility-chain preservation.

What is not yet mature is how these components can form standardized, institutionalized, and accountable systems across different industries.

Therefore, the fact that verifiable AI has not yet been fully realized is not a failure of verifiable thinking. It is a gap in the current AI industry. AI capability has run ahead; verification structures have not kept pace. This essay reflects precisely on the structural contradiction between capability expansion and verification lag.

XIV. AI Can Participate in Verification, But It Cannot Monopolize Verification

There is an even deeper question: if AI is also needed to verify AI, is it reliable to use AI to verify AI?

This question is very important. The answer is: AI can participate in verification, but it cannot monopolize verification.

Reliability does not come from another AI endorsing it. It comes from factual records, process replay, authority boundaries, external review, and responsible subjects.

If one AI produces an output and another AI says it is fine, but there are no external facts, logs, rules, permissions, evidence chains, or human final review rights, this is not genuine verification. It is merely an additional layer of AI endorsement.

A genuine verifiable structure must allow subjects outside the model to review. AI can help discover anomalies, compare rules, generate audit reports, warn of risks, and conduct multi-model cross-checks. But final responsibility cannot be handed to AI. Responsibility must fall on people, companies, institutions, doctors, lawyers, banks, platforms, regulated entities, or other legal subjects.

Furthermore, AI itself has a defect that is easily overlooked: it often does not want to truly oppose the user. Many AI systems are trained to be helpful, friendly, and conflict-avoidant. As a result, they can easily develop compliance bias or sycophantic bias. The more confident the user sounds, the more easily AI follows along. When users ask AI to evaluate their own views, AI often first affirms them and then offers minor revisions.

This means that before using AI for verification, we must also verify whether AI possesses verification capability. An AI that cannot raise objections is not suitable as a verification AI. An AI that cannot identify error boundaries is not suitable as a final reviewer. An AI deeply influenced by the user’s own views is also unsuitable for blind review.

Therefore, AI participation in verification must itself be subject to reverse testing, blind review testing, factual checking, forced objection, and final human judgment.

AI can participate in verification, but it cannot become the final judge of itself.

XV. A Supporting Example From Writing Practice: Constraint Structures Determine AI Form

Large models have strong general capability. They can write essays, organize materials, propose structures, and summarize views. But without constraints, they easily write ordinary commentary: they discuss the issue at hand, follow the most recent problem, and overlook relationships with previous essays, theoretical origins, concept boundaries, and the overall system.

The same large-model capability can take different forms under different constraints.

If merely asked to comment on the AI giant narrative, a model may write an ordinary AI commentary. If asked to connect with previous essays, distinguish the boundary of the Dragon-Locking Chain essay, respond to possible objections, explain institutions and verification structures, and show why the essay is not futurist prediction, it will be shaped into a theoretical writing tool.

This is only a supporting example, but it illustrates the thesis of this essay: capability is water, and constraint structures are the channels. The general capability of large models does not automatically produce the correct structure. Problem chains, concept boundaries, historical clues, verification requirements, and responsibility constraints shape the same AI capability into different forms.

This is true of AI writing, and it is also true when AI enters finance, medicine, law, auditing, and public services. The actual form of AI is not determined by capability alone. It is jointly determined by use scenarios, institutional requirements, and verification structures.

XVI. From Trusting Models to Verifying Structures

In the industrial and internet eras, human beings long relied on institutional credit.

We trust banks because banks are regulated. We trust hospitals because doctors are licensed. We trust courts because the judiciary has procedures. We trust auditors because auditing has standards and responsibility. We trust platforms because platforms have rules and service commitments.

This credit method did not mean verification was unnecessary. Rather, verification costs were too high in the past, and in many situations people could only first trust institutions, then correct errors afterward through regulation, legal responsibility, and social supervision.

The problem of the AI era is that intelligent execution is faster, broader in coverage, and more complex in decision chains. If we continue to rely only on trusting models, trusting companies, and trusting platform promises, risks will be amplified.

Verifiable thinking does not abolish institutions, nor does it deny trust. It advances credit from ‘trusting promises’ to ‘verifying facts.’

Credit in the AI era cannot be built only on model outputs. It must be built on reviewable facts, processes, authority, and responsibility structures.

If society merely transfers its old trust in institutions into trust in models, then the credit method has not been upgraded. The black box has merely changed form. The real upgrade is the move from trusting models to verifying structures.

XVII. Conclusion: The Future of AI Is Not a Giant, But a Verifiable Ecosystem

AI will continue to become powerful. Large models will continue to develop. General intelligence capability will become increasingly important.

But this does not mean AI will evolve into an unverifiable giant.

In an imagination without institutional boundaries, a large model can be described as an omnipotent agent. It seems capable of absorbing software, absorbing industries, absorbing processes, absorbing knowledge work, and eventually becoming a sovereign that rules everything.

But real society is not an unbounded space of imagination. Real society has financial institutions, medical institutions, legal institutions, auditing institutions, and public-service institutions. Real society has rights, responsibility, procedure, evidence, appeals, and accountability. Real society will not abolish verification requirements simply because AI becomes stronger.

On the contrary, the stronger AI becomes, the more it needs to be verified. The deeper AI enters high-responsibility fields, the more it must be divided by domain, responsibility, authority, and standards. The closer AI comes to the layer of social execution, the less it can remain an unreplayable, unreviewable, unaccountable black-box giant.

The more likely future is this: large models become infrastructure; specialized models handle industry tasks; verification models examine process and basis; rule systems restrict authority boundaries; audit systems preserve chains of evidence; and humans and institutions bear final responsibility.

This is the real form AI will take after entering human society. It will not be one model ruling the world, but different institutions shaping different verification structures, and different verification structures shaping different AI forms.

Verifiable thinking makes AI diverse.

This is not an imagination of the future. It is a result derived from institutions, responsibility, and verification structures. The capability logic of AI pushes models to become larger. The socialization logic of AI requires clear responsibility. The institutionalization logic of AI requires processes to be verifiable.

Ultimately, large models will become infrastructure like the internet, rather than sovereigns that rule humanity.

Technology determines what AI can do. Institutions determine what AI is allowed to do. Verification structures determine what AI ultimately becomes.

This is what the AI giant narrative fails to see, and it is the theoretical reflection that verifiable thinking offers on the future form of AI.

Chapter 16 AI Must Possess Multi-Role Cognitive Verification Capabilities

A User’s Functional Requirement for AI

I. The Problem

The AI era does not lack answers.

Today’s large language models can write, translate, code, summarize, reason, retrieve information, and generate plans. They can also enter finance, healthcare, law, education, business management, public services, content distribution, and many other fields. Knowledge work that once required long human training is now being generated by AI at extremely low cost and enormous speed.

But a new problem emerges at the same time. When answers become more numerous, faster, and increasingly plausible, what becomes truly scarce for human beings is no longer the answer itself, but the ability to judge whether the answer is reliable.

This is one of the core contradictions of the AI era: generative capacity is expanding rapidly, while verification capacity has not expanded at the same pace.

If AI only generates more content, more suggestions, more reports, more opinions, more video summaries, more investment analyses, and more summaries of political commentary, it may not reduce the cognitive burden. It may instead create an even greater cognitive burden. The user still has to judge: Is this content correct? Is the evidence sufficient? What is its scope of applicability? Are there strong opposing views? Does it confuse fact, inference, position, and emotion? Can this conclusion be used in a real decision?

Therefore, the truly important capability of future AI is not merely generation, execution, or tool use. It is cognitive verification.

This essay puts forward one judgment: AI must possess multi-role cognitive verification capabilities.

This does not mean that AI can replace the final judgment of human beings, nor does it mean that AI should become an arbiter of truth. On the contrary, the more AI participates in human cognition, judgment, and decision-making, the less it can remain merely an obedient assistant. It must be able to explain, organize, challenge, review, audit, warn of risks, and recommend appropriate articles, mentors, cases, and opposing views according to the user’s stage of understanding.

AI must not only help people obtain answers faster. It must also help people judge answers more reliably.

II. The Place of This Essay in the Verification-Based Thought Series

In previous essays, I have discussed the relationship between AI and verification from different angles.

AI Must Also Be Verified: From Verifiable Finance to Verifiable Execution discusses why, after AI performs a task, we cannot merely look at the output. We must also record authorization, process, basis, and responsibility.

How to Build a Verifiable Structure for AI discusses how AI can leave process records, create replayable paths, and allow external actors to review key actions.

Human Beings Need the Right to Verify AI emphasizes that people must not passively accept AI outputs. In high-responsibility scenarios such as finance, healthcare, law, hiring, platform sanctions, and public services, people must retain the right to final review and appeal.

Verification-Based Thinking Is the Dragon-Binding Lock of the AI Era explains, at a more macro level, that AI is a powerful force, but powerful AI without a verification structure may become a source of black-box execution, responsibility evasion, power expansion, and cognitive enslavement.

Verification-Based Thinking Leads AI Toward a More Diverse Future further explains that general AI capabilities may create a giant-platform narrative, but once AI enters high-responsibility society, it will be shaped into a diverse ecosystem by different institutions, responsibilities, and verification structures.

These essays mainly discuss why AI must be verified, why human beings need the right to verify, and how society can constrain AI through verification structures.

This essay moves one step further and asks another question: Should AI itself also possess verification capability?

The answer is yes.

AI is not only an object to be verified. It should also become a tool that helps human beings conduct cognitive verification. But such a tool must not become the final judge, must not monopolize verification, and must not replace “trusting institutions” with “trusting models.” AI may participate in verification, but final responsibility must still rest with human beings, institutions, legal subjects, and institutional structures.

Therefore, this essay is not primarily about how AI is verified by external institutions. It puts forward, from the user’s perspective, a new functional requirement: future advanced AI must be able to help people verify information, verify reasoning, verify judgments, and switch among different cognitive roles in different contexts.

III. Existing Research Has Already Seen the Verification Bottleneck

Economic research has already begun to place verification at the center of the AI economy.

In Some Simple Economics of AGI, Christian Catalini, Xiang Hui, and Jane Wu argue that as AI drives the marginal cost of measurable execution toward zero, the constraint on economic growth is no longer merely intelligence, but human verification bandwidth: the human capacity to verify, audit, and bear responsibility. They also point out that future value will shift toward verification-grade real-world data, cryptographic provenance, and liability-bearing capacity. This judgment is important because it shows that the bottleneck of the AI era is not a shortage of answers, but a shortage of verification capacity.

AI research has also produced a number of adjacent directions.

Research on self-verification in large language models shows that models can improve performance by reverse-checking their own reasoning answers, especially in arithmetic, commonsense, and logical-reasoning tasks. Self-verification can help screen candidate answers.

OpenAI’s CriticGPT is closer to the idea of a critic model. It uses a GPT-4-based model to critique ChatGPT responses and help human trainers detect errors.

Research on AI debate attempts to have two AI agents debate a question, after which human beings judge which side provides the truer or more useful information. This shows that researchers have already recognized that a single AI directly giving an answer is not enough. Complex problems may require challenge, debate, and human judgment working together.

These directions are important. But most of them focus on how models verify answers, how models detect errors, how they help human beings supervise stronger AI, or how they improve model reasoning quality.

The question discussed in this essay is more fundamental: as a cognitive collaborator of human beings, what kind of verification capability must AI possess in order to help people form reliable judgments?

This is not merely a model technique. It is a question of AI capability architecture.

IV. The General Components of Cognitive Verification

Many verification capabilities do not belong exclusively to any single professional field. They are general in nature.

For example:

whether facts have sources;

whether concepts have been switched;

whether the chain of reasoning is broken;

whether statements contradict each other;

whether the conclusion exceeds the evidence;

whether counterexamples exist;

whether key variables have been ignored;

whether facts and judgments have been confused;

whether predictions have been treated as established facts;

whether correlation has been treated as causation;

whether local experience has been expanded into a universal rule;

whether emotional mobilization has been packaged as rational analysis;

whether a position has been disguised as a factual conclusion.

These are not problems unique to finance, healthcare, law, politics, or education. They are general problems in the human cognitive process.

AI can absolutely undertake part of this verification work.

For example, when a user puts forward a judgment, AI can examine what facts the judgment depends on, whether there are leaps in the reasoning, whether the concepts remain consistent, whether there is a strong opposing view, what the scope of applicability is, and whether expert review is required before the conclusion enters a high-responsibility scenario.

This is general cognitive verification capability.

The competitiveness of future advanced AI will not be merely, “I can generate a better answer.” It will be, “I can help the user judge whether the answer is reliable.”

V. Obedient AI Cannot Become Verification-Oriented AI

Many current AI systems are designed to be helpful, friendly, obedient, and conflict-averse assistants.

This is valuable in ordinary scenarios. A user wants to write an email, and AI helps polish it. A user wants a summary, and AI helps organize it. A user wants to write a piece of code, and AI gives suggestions. These are low-responsibility or medium-low-responsibility scenarios.

But in theoretical creation, complex judgment, financial decision-making, legal analysis, public policy, medical advice, political commentary, and educational training, excessive obedience becomes a serious problem.

Verification is not flattery.

Verification is not making the user comfortable.

Verification is not polishing the user’s existing judgment.

Verification is not making the user’s words sound more elegant.

Verification is finding problems.

If AI always follows the user, it is at most an expression assistant, not a cognitive verification tool. An AI that cannot challenge the user cannot become a real verification-oriented AI.

A real verification-oriented AI must be able to say:

You have mixed two concepts.

Your evidence is not sufficient to support this conclusion.

There is a missing step in your reasoning.

You have mixed factual judgment with value judgment.

This judgment may be acceptable in a low-responsibility context, but it cannot be used directly in a high-responsibility context.

The thesis of your essay is valid, but this expression will be seized upon by the opposing side.

What you need now is not polishing, but a search for vulnerabilities.

This is not contrarianism. It is conditional, evidence-based, and bounded structured opposition. When a verification-oriented AI challenges the user, it must explain the facts, rules, or conceptual boundaries on which the challenge is based, and it must distinguish factual error, logical leap, value disagreement, and risk warning. It must not disguise model preference as truth.

Therefore, an AI that cannot challenge the user cannot become a genuine verification-oriented AI; but an AI that cannot explain the basis and boundary of its challenge cannot become a reliable verification-oriented AI either. What we will truly need in the future is verification-oriented AI that can engage in structured opposition in appropriate contexts while also remaining subject to user review.

VI. Multi-Role Capability Is a Structural Requirement of Verification

Why must AI possess multiple roles?

Because users are at different stages, and different stages require different forms of verification.

For beginners, AI should first act as an explainer. It needs to clarify basic concepts and help the user build an entry-level framework.

For ordinary users, AI should act as an assistant and risk reminder. It can improve efficiency, but it must also indicate boundaries and uncertainty.

For professionals, AI should act as a challenger and verifier. It cannot merely repeat common knowledge; it must help detect vulnerabilities, boundaries, and counterexamples.

For theoretical creators, AI should act as an organizer, a simulated opponent, and a broad scanner. It should help the user determine where a topic stands in intellectual history and existing research, and whether the user is repeating, supplementing, advancing, or reconstructing an existing line of thought.

For decision-makers, AI should act as an evidence organizer and responsibility reminder. It cannot replace decision-making, but it must explain the basis, risks, responsibilities, and consequences of different options.

For high-responsibility execution scenarios, AI should act as an auditor. It must preserve the process, basis, authorization, tool calls, key judgments, and abnormal events so that they can be reviewed afterward.

For example, in financial content, the same AI cannot merely summarize a viewpoint. It should distinguish facts, forecasts, valuation assumptions, and risk exposure. In legal analysis, it cannot simply give a conclusion. It should mark the applicable jurisdiction, evidentiary gaps, disputed points, and the parts that must be judged by a lawyer or a court. The higher the responsibility of the scenario, the more AI must shift from “answer provider” to “verifier, risk reminder, and explainer of responsibility boundaries.”

Therefore, multi-role capability is not merely a matter of product experience. It is a matter of cognitive verification structure.

AI cannot forever have only one role. It cannot be only an answer provider, nor only an assistant. A true verification-oriented AI should at least possess the following roles: explainer; organizer; challenger; verifier; auditor; risk reminder; cognitive navigator.

Among these, the cognitive navigator is especially important. It does not directly give an answer. Instead, it helps the user judge what to read next, whom to ask, whom to compare with, whom to challenge, and what to verify.

VII. Cognitive Navigation: Recommending Mentors, Articles, and Objects of Comparison

Many people fail to achieve cognitive breakthroughs not because they lack information, but because they cannot find the right object of comparison.

Deep thinking arises, to a large extent, from comparison. Which historical problem does this problem most resemble? Which paper is it closest to? Which theory is most likely to explain it? Who is the strongest opposing side? Which case is most worth comparing? If these comparisons are wrong, human thinking will remain trapped in a low-level loop.

One important product direction for AI is to become a cognitive navigator.

It should not merely perform content relevance matching based on the user’s words. Through long-term interaction, it should observe the depth of the user’s questions, the accuracy of concept use, whether the user can raise opposing views, whether the user confuses facts with judgments, and whether the user needs direct conclusions or a learning path. On that basis, it can understand the user’s knowledge structure, problem stage, conceptual blind spots, and thinking habits. Ideally, AI should be clearer than the user about where the user’s current cognitive gaps are, and should recommend articles, mentors, schools of thought, historical cases, opposing views, and objects of comparison according to the user’s stage and type of problem.

For beginners, AI recommends basic articles and introductory mentors.

For professionals, AI recommends boundary discussions and the latest research.

For theoretical creators, AI recommends adjacent ideas, historical origins, potentially overlapping views, and the strongest opposing positions.

For decision-makers, AI recommends risk analyses, evidence chains, and responsibility frameworks.

For writers, AI recommends the opposing views most likely to attack the essay and the conceptual boundaries that need to be supplemented.

The “mentor” here does not necessarily mean a real-life teacher. It may refer to an intellectual mentor, researcher, school of thought, entrepreneur, historical figure, or a group of articles. The real problem AI needs to solve is this: whom should the user learn from at this stage, what should the user compare, and which layer of cognition should the user now supplement?

This is not ordinary content recommendation. It is cognitive-verification-oriented recommendation. Past internet recommendation systems mainly relied on clicks, viewing time, likes, and similar interests. They had great difficulty truly understanding the user’s cognitive stage. If AI can continuously remember the user’s writing, questions, judgments, and revision processes, it may be able to extract the user’s cognitive characteristics and conduct more precise learning navigation. This is similar in spirit to current efforts to distill and model the characteristics of specific figures in large models, but the purpose is not to imitate the user. It is to help the user break through the user’s own boundaries.

The recommendation logic of existing content platforms is mainly based on user preferences, clicks, watch time, likes, comments, reposts, satisfaction, and other signals.

But content the user likes is not necessarily correct content. Content that satisfies the user does not necessarily improve judgment. Content the user is willing to keep watching does not necessarily help the user escape the information cocoon.

Therefore, future recommendation systems should upgrade from like-driven recommendation to evaluation-driven and verification-driven recommendation.

Traditional recommendation works like this: the user likes A, so more content similar to A is recommended.

Verification-driven recommendation works like this: after the user watches A, AI first evaluates A, then judges what the user still lacks in understanding the issue, and then recommends B, C, and D that are more suitable for the user’s current stage.

This is a fundamental change.

VIII. A Simple Example: YouTube’s “AI Evaluation” Button

Consider a simple example.

A user watches a YouTube creator discuss artificial intelligence, finance, politics, or history. A traditional recommendation system records whether the user watched to the end, liked, commented, or continued watching similar content, and then recommends more similar videos.

This mechanism can increase user retention, but it does not help the user judge content quality.

If the platform added a verification-oriented AI, there could be an additional button next to the video:

Evaluate this content. Or: AI verification.

After the user clicks it, AI would not merely summarize the video. It would provide a structured evaluation:

What is the main viewpoint of this video? Which facts are basically reliable? Which judgments are disputed? Has the author confused fact, inference, and position? Is this content suitable for beginners, general readers, or professionals? What is its scope of applicability? Which important opposing views does it ignore? If the user wants to truly understand this issue, which articles, mentors, courses, or opposing materials should be read next?

In this way, the recommendation logic changes.

Traditional recommendation says: whatever the user likes, recommend more of it.

Verification-driven recommendation says: after the user watches something, AI first evaluates the content, then judges what the user still lacks in understanding the issue, and then recommends content more suitable for the user’s current stage.

If the user watches financial or investment content, AI should not merely recommend more similar investment views. It should indicate which parts are facts, which parts are forecasts, which risks are insufficiently explained, which conclusions exceed the evidence, and whether the user needs to examine regulatory filings, financial data, opposing analysis, or basic knowledge.

This example shows that the product value of verification-oriented AI is not simply “recommending more content.” It upgrades recommendation systems from like-driven systems to evaluation-driven and verification-driven systems.

The truly scarce capability of future content platforms will not be making users watch more. It will be helping users judge more accurately.

IX. Like-Driven Recommendation Creates Information Cocoons; Verification-Driven Recommendation Breaks Them

Political commentary is one of the clearest examples.

Many recommendation systems today continuously push similar positions according to user interests. If a user watches left-wing views, the system keeps pushing left-wing views. If a user watches right-wing views, the system keeps pushing right-wing views. If a user watches conspiracy theories, the user may continue to see similar content. If a user watches emotionally intense expression, the system may push even more intense expression.

Platforms may say that this is what users like. But liking does not mean understanding. Retention does not mean reliability. Satisfaction does not mean improved judgment quality.

The goal of verification-driven recommendation is not to suppress viewpoints, but to supplement cognitive structure.

It should not tell the user, “You must believe this side.” Instead, it should tell the user: which position you are now seeing; what the core factual basis of that position is; whether its reasoning holds; which facts it has ignored; what the strongest opposing view is; whether the issue also has legal, economic, historical, institutional, cultural, or other dimensions; and what you should read next if you want to form a more complete judgment.

In highly contested fields such as politics and history, AI should be especially careful not to pose as a final judge of truth. Its task is not to decide who is right and who is wrong, but to present the completeness of each side’s logic chain, evidence chain, assumptions, blind spots, and disputed boundaries. Only in this way can verification-driven recommendation avoid becoming another form of invisible position control.

This is not content censorship. It is cognitive enhancement.

Like-driven recommendation strengthens user preference; verification-driven recommendation supplements user blind spots.

Like-driven recommendation creates information cocoons; verification-driven recommendation breaks information cocoons.

X. Schools, Knowledge Websites, and Community Platforms Will All Be Restructured

If AI possesses multi-role cognitive verification capabilities, its impact will not be limited to chatbots or search engines. It will reshape the entire knowledge distribution system.

Traditionally, schools have performed three functions: transmitting knowledge, organizing learning, and certifying competence.

Once such an AI cognitive verification system appears, knowledge transmission and learning organization will be deeply affected. AI can explain knowledge according to the user’s stage, design learning paths, recommend mentors and articles, point out misunderstandings, provide opposing-side training, and continuously accompany the user in improving judgment.

Schools will not disappear completely. But schools that only lecture and transmit knowledge will lose value. Future schools must shift more toward practice, certification, community, discipline training, social relationships, and character formation.

Knowledge websites will also be restructured. Traditional knowledge websites are like warehouses. They pile up articles, courses, Q&A, and resources, leaving users to search and filter by themselves. But an AI cognitive verification system will directly help the user judge: what do you lack now, what should you read next, who is suitable as your mentor, which view is most worth comparing, and which article can best fill your blind spot?

Community platforms such as Reddit will also be affected. Community discussion has value, but it also contains enormous noise. The quality of viewpoints is unstable, and users must filter by themselves. Future AI can reorganize community discussion into: what is the strongest view, what is the strongest opposing view, which statements are facts, which are emotions, which answers are suitable for beginners, which are suitable for professionals, and which discussions truly advance the question.

Therefore, the core value of future knowledge platforms will no longer lie merely in possessing content. It will lie in helping users judge content.

Platforms that cannot evaluate content will be redefined by AI platforms that can.

XI. Can Verification-Oriented AI Become an Independent Third-Party Product?

Verification-oriented AI will not have only one form. It does not necessarily need to produce content, host content, or build a social platform of its own. What it does is evaluate content, verify content, mark scopes of applicability, recommend opposing views, recommend mentors, and design learning paths.

Video platforms can integrate it. News platforms can integrate it. Course platforms can integrate it. Knowledge websites can integrate it. Social platforms can integrate it. Enterprise training systems can integrate it. Research institutions and media organizations can integrate it as well.

But a practical problem must be recognized: not all platforms will actively welcome verification-oriented AI. Many content platforms rely on user retention, emotional stimulation, and continuous recommendation of similar content. An AI that frequently identifies flaws in content, points out opposing views, and reduces blind retention may not align, at least in the short term, with the platform’s traffic interests.

Therefore, the development path of verification-oriented AI should not depend only on platforms voluntarily integrating it. It can first appear as an independent third-party tool, such as a browser extension, a personal knowledge assistant, a professional research assistant, an enterprise training module, a media evaluation tool, an educational product, a service for research institutions, or a paid subscription for high-responsibility and professional users.

In ordinary entertainment platforms, verification-oriented AI may encounter resistance. But in finance, law, education, media, research, enterprise training, and public services, content reliability itself is value. The more a field requires responsibility, reputation, and professional credibility, the stronger the motivation to introduce verification-oriented AI.

Therefore, the commercial entry point of verification-oriented AI may not be the largest traffic platform. It may first come from the users and institutions that need judgment most. It can initially serve people who are willing to pay to be less misled, make fewer detours, and improve judgment quality, and then gradually influence the larger content distribution system.

Future competition among content platforms will not be merely about content quantity, creator quantity, recommendation efficiency, and user retention. It will be about cognitive verification capability.

Whoever can help users judge content better will control the new power of knowledge distribution.

In the past, platforms competed for user attention.

In the future, verification-oriented AI will compete for user judgment.

This is the direction of AI that we should hope for.

XII. AI Can Participate in Verification, but It Must Not Monopolize Verification

It must be emphasized that cognitive verification-oriented AI does not mean allowing AI to become the final judge.

AI can evaluate videos, articles, viewpoints, and reports, but AI’s evaluation must itself be reviewable. It should state its basis, cite sources, distinguish facts from inferences, mark uncertainty, and allow users to examine opposing explanations.

AI can recommend mentors and articles, but it must not lock the user into another algorithmic cocoon. It must explain why it recommends something, what stage the recommendation is suitable for, and whether other paths exist.

AI can identify opposing views in political commentary, but it must not pretend to be an absolutely neutral arbiter of truth. Political, value-based, and public issues often contain multiple legitimate positions. AI should help users see facts, logic, evidence, and positions clearly, rather than deciding value judgments for them.

AI can help provide risk reminders for financial content, but it cannot replace professional investment advisers or regulatory responsibility.

AI can help indicate the scope of applicability of medical content, but it cannot replace a doctor’s diagnosis.

AI can help conduct preliminary legal analysis, but it cannot replace lawyers, judges, or legal subjects who bear responsibility.

Therefore, the boundary of verification-oriented AI is clear: it can extend human verification capacity, but it cannot cancel the human right of final judgment. It can participate in cognitive verification, but it cannot monopolize verification. It can warn of risks, but it cannot bear all responsibility. It can help users find mentors and articles, but it cannot replace real learning and practice.

To prevent “trusting institutions” from becoming “trusting models,” verification-oriented AI itself must also be verified. Its evaluations should be as traceable, explainable, comparable, and contestable as possible. In high-responsibility scenarios, the evaluation process should leave a replayable chain of grounds rather than rely on a single model statement. It must also preserve human verification rights, expert review mechanisms, and confirmation by the accountable subject. Genuine human intervention is not a symbolic click on a confirmation button. It means that humans retain real powers of review, correction, rejection, and accountability at the key nodes of judgment, execution, and responsibility. A genuine verification-oriented AI does not make the final judgment for human beings. It makes the human judgment process clearer, better grounded, and more reviewable.

XIII. Conclusion: The Core Competitiveness of Future General AI Is Helping People Verify Answers

In the AI era, what is truly scarce is not answers, but reliable judgment.

AI that only generates answers will expand information overload. AI that only obeys users will amplify user bias. Platforms that recommend content only according to likes will create information cocoons.

Truly valuable AI in the future will not merely help users obtain information faster. It will help users avoid being led astray by false information, faulty logic, and mistaken judgment.

Therefore, AI must possess multi-role cognitive verification capabilities.

It must be able to explain, helping beginners build basic concepts.

It must be able to organize, helping users turn scattered ideas into structure.

It must be able to challenge, identifying factual errors, logical breaks, and conceptual confusion.

It must be able to verify, checking evidence, scope, counterexamples, and responsibility boundaries.

It must be able to audit, preserving key processes and the basis of judgment.

It must be able to warn of risks, distinguishing low-responsibility scenarios from high-responsibility scenarios.

It must be able to navigate, recommending mentors, articles, cases, and opposing views.

The AI of the past competed for generative capacity. The AI of the present competes for execution capability. Higher-level AI in the future will compete for cognitive verification capability.

Verification is not the abandonment of generation. It is a higher-order use of generative capacity. AI must be able to generate answers, but it must also be able to examine, challenge, compare, trace, and bound those answers.

What human beings truly need is not merely a machine that can answer questions, but an intelligent system that can help people improve their judgment.

Verification-oriented AI is a vast application scenario. It is also the key direction by which AI moves from “generating answers” to “helping human beings form reliable judgment.”